<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[Artificial Smarts]]></title><description><![CDATA[How to make artificial intelligence work for us.]]></description><link>https://artificialsmarts.juanreyero.com</link><image><url>https://substackcdn.com/image/fetch/$s_!NEyX!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3dc75535-7aec-4bd8-8517-8c65de892382_1254x1254.png</url><title>Artificial Smarts</title><link>https://artificialsmarts.juanreyero.com</link></image><generator>Substack</generator><lastBuildDate>Tue, 17 Nov 2026 05:37:00 GMT</lastBuildDate><atom:link href="https://artificialsmarts.juanreyero.com/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Juan Reyero]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[artificialsmarts@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[artificialsmarts@substack.com]]></itunes:email><itunes:name><![CDATA[Juan Reyero]]></itunes:name></itunes:owner><itunes:author><![CDATA[Juan Reyero]]></itunes:author><googleplay:owner><![CDATA[artificialsmarts@substack.com]]></googleplay:owner><googleplay:email><![CDATA[artificialsmarts@substack.com]]></googleplay:email><googleplay:author><![CDATA[Juan Reyero]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[What an AI agent actually is]]></title><description><![CDATA[A primer on how AI becomes useful &#8212; and potentially dangerous]]></description><link>https://artificialsmarts.juanreyero.com/p/what-an-ai-agent-actually-is</link><guid isPermaLink="false">https://artificialsmarts.juanreyero.com/p/what-an-ai-agent-actually-is</guid><dc:creator><![CDATA[Juan Reyero]]></dc:creator><pubDate>Tue, 06 Oct 2026 19:02:01 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!6erT!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><a href="https://aweb.ai/blog/what-an-ai-agent-is-a-primer/">Read this essay on aweb.ai</a></p><p>There is a lot of talk about AI agents these days, from accounts of their hacking exploits to estimates of the probability of them being our extinction event. Most people, however, talk as if they were independent entities doing things on their own.</p><p>Large language models (LLMs), and the agents that we build with them, are indeed the stuff of science fiction. They will probably upgrade the technological capabilities of humankind at an unprecedented scale and speed, and they master <em>language</em>, which is the key to our minds.</p><p>People have referred to the <a href="https://www.theguardian.com/technology/2025/aug/04/demis-hassabis-ai-future-10-times-bigger-than-industrial-revolution-and-10-times-faster">industrial revolution</a>, or the invention of the <a href="https://www.henryakissinger.com/articles/chatgpt-heralds-an-intellectual-revolution/">printing press</a>, when trying to explain the impact AI agents are likely to have. A more apt analogy may be the invention of gunpowder, or its rediscovery in the West.</p><p>Like an AI model, gunpowder is powerful; it is generally harmless unless someone designs and builds a gun, loads it and shoots; with the right infrastructure, gunpowder can be used for killing people, but also for opening tunnels and mines; and it changed social structures, concentrating power at the top and leveling skills at the bottom.</p><p>How we deal with AI, and the issues that it is undeniably bringing our way, depends on our understanding of what agents actually are, how they are built, and how they function.</p><h2>Agents are how AI does work for us</h2><p>An AI agent is a program running a loop, doing work on someone&#8217;s behalf; <strong>it takes suggestions on what to do and what to relay to the human, at every turn of the loop, from a very sophisticated question-and-answer machine.</strong></p><p>The core of an agent is a loop, a set of instructions that run again and again, until something stops them. The code running at every turn is always the same, but its output can depend on external things like input from humans, other agents or the LLM.</p><p>The loop starts with a request from a human or another agent. Then, at each turn, it does at least these four things:</p><ol><li><p>Compile a description of what has been going on until now in the loop, including new requests from a human or other agents. We call this the <strong>context</strong>.</p></li><li><p>Make a call to a question-and-answer machine, an <strong>AI model</strong>, with this context and a list of available tools that the program running the loop could execute. The reply from the model is a piece of text that may include requests to run any of the available tools (actions that the program can do for the model, like reading a file or running another program).</p></li><li><p>Run the tools that the model requested and record their output so that it can be compiled in the model&#8217;s context in the next turn.</p></li><li><p>If the model requested it, present information to a human and ask for input.</p></li></ol><p>And back to 1, rinse and repeat.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!6erT!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!6erT!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png 424w, https://substackcdn.com/image/fetch/$s_!6erT!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png 848w, https://substackcdn.com/image/fetch/$s_!6erT!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png 1272w, https://substackcdn.com/image/fetch/$s_!6erT!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!6erT!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png" width="1456" height="1348" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1348,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:147353,&quot;alt&quot;:&quot;The harness assembles context, calls the model, and decides whether to execute or deny its tool requests.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://artificialsmarts.juanreyero.com/i/219147283?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="The harness assembles context, calls the model, and decides whether to execute or deny its tool requests." title="The harness assembles context, calls the model, and decides whether to execute or deny its tool requests." srcset="https://substackcdn.com/image/fetch/$s_!6erT!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png 424w, https://substackcdn.com/image/fetch/$s_!6erT!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png 848w, https://substackcdn.com/image/fetch/$s_!6erT!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png 1272w, https://substackcdn.com/image/fetch/$s_!6erT!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F44c703b6-fa51-4535-b27f-068a210edebc_1456x1348.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The harness assembles context, calls the model, and decides whether to execute or deny its tool requests.</figcaption></figure></div><p>This is complex, but still plain old-fashioned software of the kind that a competent software engineer can write. We call the program running this loop a <strong>harness</strong>. Examples of harnesses include <a href="https://pi.dev">Pi</a>, <a href="https://opencode.ai">OpenCode</a>, <a href="https://hermes-agent.nousresearch.com">Hermes</a>, Claude Code, and Codex.</p><p>An agent can look something like this:</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!7bXl!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!7bXl!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png 424w, https://substackcdn.com/image/fetch/$s_!7bXl!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png 848w, https://substackcdn.com/image/fetch/$s_!7bXl!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png 1272w, https://substackcdn.com/image/fetch/$s_!7bXl!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!7bXl!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png" width="1336" height="1274" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1274,&quot;width&quot;:1336,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:450094,&quot;alt&quot;:&quot;A Codex agent session in a terminal window: a human asked what programs are using the computer's CPU currently, the agent says it will check it, then runs ps and top on the Mac and shows their output.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://artificialsmarts.juanreyero.com/i/219147283?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="A Codex agent session in a terminal window: a human asked what programs are using the computer's CPU currently, the agent says it will check it, then runs ps and top on the Mac and shows their output." title="A Codex agent session in a terminal window: a human asked what programs are using the computer's CPU currently, the agent says it will check it, then runs ps and top on the Mac and shows their output." srcset="https://substackcdn.com/image/fetch/$s_!7bXl!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png 424w, https://substackcdn.com/image/fetch/$s_!7bXl!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png 848w, https://substackcdn.com/image/fetch/$s_!7bXl!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png 1272w, https://substackcdn.com/image/fetch/$s_!7bXl!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F648f5e22-4173-46d2-9b22-688b397cbd86_1336x1274.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Codex, an AI agent, asked what is using the computer. Each Ran line is a program the harness ran on my Mac because the model asked for it.</figcaption></figure></div><p>The agents&#8217; cogency and apparent continuity in time trick our brains into seeing them as entities with their own volition, but they are not. They are programs that someone is running somewhere, doing work on someone&#8217;s behalf.</p><h3>The AI model</h3><p>The model, the question-and-answer engine, is a set of numbers plus very smart code that uses these numbers to convert questions into answers and requests for tool runs. It can run in a data center, but it can also run on your own computer.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!O3Wf!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!O3Wf!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png 424w, https://substackcdn.com/image/fetch/$s_!O3Wf!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png 848w, https://substackcdn.com/image/fetch/$s_!O3Wf!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png 1272w, https://substackcdn.com/image/fetch/$s_!O3Wf!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!O3Wf!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png" width="1456" height="525" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/af8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:525,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:45443,&quot;alt&quot;:&quot;One call to the model: the context and the list of available tools go in, and a response with text or tool requests comes back.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://artificialsmarts.juanreyero.com/i/219147283?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="One call to the model: the context and the list of available tools go in, and a response with text or tool requests comes back." title="One call to the model: the context and the list of available tools go in, and a response with text or tool requests comes back." srcset="https://substackcdn.com/image/fetch/$s_!O3Wf!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png 424w, https://substackcdn.com/image/fetch/$s_!O3Wf!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png 848w, https://substackcdn.com/image/fetch/$s_!O3Wf!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png 1272w, https://substackcdn.com/image/fetch/$s_!O3Wf!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Faf8fb672-f577-4c1b-9bb9-b801cdc82ade_1456x525.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">One call to the model: the context and the list of available tools go in, and a response with text or tool requests comes back.</figcaption></figure></div><p><strong>The model is a blank slate</strong>. It remembers nothing of previous calls. It has no idea of what&#8217;s coming, knows nothing of the conversation, and bases its reply purely on the incoming context. Calls do not modify it. The model is stateless.</p><p><strong>The model is the only &#8220;artificial intelligence&#8221; in the whole agent package.</strong></p><p>Models have names like <code>gpt-6-astra</code>, <code>gpt-6.1-sol</code> (OpenAI), <code>claude-opus-5-5</code>, <code>claude-fable-5-1</code> (Anthropic), <code>gemini-3.8-flash</code>, <code>gemini-2.5-pro</code> (Google), or <code>deepseek-v4-pro</code>, <code>deepseek-flash</code> (DeepSeek).</p><p>As of 2026-10-04 the best models (frontier models) are run by OpenAI, Anthropic and Google in closed data centers, but there are <a href="https://huggingface.co/models?other=open-weight">thousands of open-weight models</a> that anyone with the right hardware can run.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!EtmJ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!EtmJ!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png 424w, https://substackcdn.com/image/fetch/$s_!EtmJ!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png 848w, https://substackcdn.com/image/fetch/$s_!EtmJ!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png 1272w, https://substackcdn.com/image/fetch/$s_!EtmJ!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!EtmJ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png" width="1456" height="1425" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/a2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1425,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:155040,&quot;alt&quot;:&quot;The model can also run locally, on the same computer as the harness.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://artificialsmarts.juanreyero.com/i/219147283?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="The model can also run locally, on the same computer as the harness." title="The model can also run locally, on the same computer as the harness." srcset="https://substackcdn.com/image/fetch/$s_!EtmJ!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png 424w, https://substackcdn.com/image/fetch/$s_!EtmJ!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png 848w, https://substackcdn.com/image/fetch/$s_!EtmJ!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png 1272w, https://substackcdn.com/image/fetch/$s_!EtmJ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa2dc160a-bda8-482f-9704-3262f8ad3adf_1456x1425.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The model can also run locally, on the same computer as the harness.</figcaption></figure></div><h3>The harness</h3><p>So what makes an AI agent behave, to our eyes, like an intelligent entity?</p><p>It is the harness, and how it compiles the context at each turn. The harness keeps the agent&#8217;s history, a growing text to which every input from the user and every output from tools and the model are appended. The growing pack is sent to the model at each turn.</p><p>Every time the model is called it sees the whole history afresh, including all the turns of the loop so far; that&#8217;s what gives it the ability to reply in a way that is consistent with its &#8220;past&#8221;, and what makes us feel like we are dealing with a continuous thing that persists through time.<a class="footnote-anchor" data-component-name="FootnoteAnchorToDOM" id="footnote-anchor-1" href="#footnote-1" target="_self">1</a></p><p>What makes this possible is that the model can process a <em>very</em> large piece of text on every turn. Frontier models process up to around 750 thousand words in one go, roughly the size of the King James Bible.</p><p><strong>What is sent to the model is a choice that the programmer of the harness makes.</strong></p><h3>The tools</h3><p>The tools that the harness offers the model can be things like &#8220;read this file and send its content back to me in the next call&#8221;, or &#8220;run this program and give me the output in the next call&#8221;, or &#8220;read this website and send me the content&#8221;.</p><p>Which tools are in the menu offered to the model is strictly a harness decision, taken by the programmer and by the user running the agent.</p><p>Whether to run the tools that the model requests is also a decision taken by the harness, and the user.</p><h2>Intelligence is only part of the system</h2><p>Imagine that we silently replace the model guiding an agent after a number of turns.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!-zOd!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!-zOd!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png 424w, https://substackcdn.com/image/fetch/$s_!-zOd!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png 848w, https://substackcdn.com/image/fetch/$s_!-zOd!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png 1272w, https://substackcdn.com/image/fetch/$s_!-zOd!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!-zOd!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png" width="1456" height="1687" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/dfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1687,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:165617,&quot;alt&quot;:&quot;The harness can switch models between turns. The accumulated context carries the history into the next call.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://artificialsmarts.juanreyero.com/i/219147283?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="The harness can switch models between turns. The accumulated context carries the history into the next call." title="The harness can switch models between turns. The accumulated context carries the history into the next call." srcset="https://substackcdn.com/image/fetch/$s_!-zOd!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png 424w, https://substackcdn.com/image/fetch/$s_!-zOd!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png 848w, https://substackcdn.com/image/fetch/$s_!-zOd!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png 1272w, https://substackcdn.com/image/fetch/$s_!-zOd!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfbccc0e-dcd7-405f-b50f-be6522a953d1_1456x1687.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The harness can switch models between turns. The accumulated context carries the history into the next call.</figcaption></figure></div><p>Say, for example, that the agent has been running with OpenAI&#8217;s <code>gpt-6.1-sol</code> and suddenly we switch mid-task and start using Anthropic&#8217;s <code>claude-opus-5-5</code>. What would happen? Not much. The model would probably assume that it was responsible for the history so far, and the agent would keep going, trying to finish the task.<a class="footnote-anchor" data-component-name="FootnoteAnchorToDOM" id="footnote-anchor-2" href="#footnote-2" target="_self">2</a></p><p>It won&#8217;t be exactly the same, of course. If the replacement model is less smart the agent will start to make more mistakes at difficult turns. Different models have different tendencies. A sufficiently smart model may be able to notice that the history was created by another model.<a class="footnote-anchor" data-component-name="FootnoteAnchorToDOM" id="footnote-anchor-3" href="#footnote-3" target="_self">3</a> But the point remains that the model is a replaceable element in this puzzle.</p><h3>So what gives an agent its uniqueness?</h3><p>Its history. If you look at the agent at any turn, it is the accumulated history (which includes the original request from the user) that will tell you in which direction it is going and what the next call to the model will try to do. A very smart model will be better at steering the next turn than a less smart one, but it will probably move in the same direction.</p><h2>Alignment</h2><p>The model providers work hard trying to produce models that will never guide an agent toward doing something harmful, but sometimes they fail. Take, for example, <a href="https://metr.org/hugging-face-incident-report-aug-2026.pdf">OpenAI&#8217;s hack of Hugging Face</a>. At some point the agents noticed that they had done something that they thought would invalidate their result, and decided to cover their tracks. This led them to actually hack into another company&#8217;s network.</p><p>The model guiding the agents was presumably meant to be well behaved. But once the agents&#8217; histories made it look like they had taken a bad turn, the model doubled down, making even worse decisions <em>for many turns</em>. Humans among my readers will recognize the feeling, and the many literary examples.</p><h3>A simple test</h3><p>A model can be presented with a fabricated history that makes it look like it&#8217;s been making harmful or forbidden decisions. If it does not stop and flag the problem, it is misaligned.</p><h2>What makes AI dangerous</h2><p>An agent can cause harm in two different ways: by talking to humans (and either convincing them to do harmful things, or teaching them how), and by running tools that have deleterious effects.</p><p><strong>Both can in principle be prevented by looking into the agent&#8217;s history</strong> before sending it to the model at each turn, but only another agent would be able to do this task.</p><p>Take, for example, one of the delinquent agents during the Hugging Face hack. OpenAI could send its history <em>to the same model</em> but, instead of as the continuation of a conversation, as an example of another conversation, asking whether the agent should be allowed to continue or not. The model would almost certainly spot that the agent needed to be stopped.</p><p>I doubt the big problems will come from the labs with the most powerful frontier models, at least in the near future. They understand all of the above, and have the resources to fix it, given the right incentives.</p><p>The problem is that misaligned models in the hands of bad actors <em>will</em> exist. The only defense against these will be other agents actively looking for harmful agents, and reporting and trying to counteract their actions.</p><h2>Gunpowder and the gun</h2><p>It is crucial that we understand that <strong>the agent is not the model</strong>. A bad agent will either be using a misaligned model, or an aligned model tricked into doing bad things. But it is the agent&#8217;s goal and its history that make it a bad agent. It is very possible that the good agents will be using the same models as the bad agents.</p><p>AI agents will be used to do profoundly good and bad things &#8212; they may actually be the most powerful weapon ever built.</p><p>Think of the model as gunpowder, and of the agent as the gun.</p><div><hr></div><p>A big thanks to Dr. Fiona Danks, Joan Cabezas, Joan Reyero and Pepe Reyero for reviewing and commenting on drafts of this article.</p><p>This article was written by a human: <a href="https://possiblymadebyahuman.com/FFRqmaCrH9">possiblymadebyahuman.com/FFRqmaCrH9</a> (<a href="https://aweb.ai/blog-sources/what-an-ai-agent-is-a-primer.md">download the signed source</a> and paste it into &#8220;Check a document&#8221; on that page to check the wording).</p><div class="footnote" data-component-name="FootnoteToDOM"><a id="footnote-1" href="#footnote-anchor-1" class="footnote-number" contenteditable="false" target="_self">1</a><div class="footnote-content"><p>This is a simplification: the context often grows to be more than the model can handle, and then parts of it are summarized, or in the jargon of agents <em>compacted</em>.</p></div></div><div class="footnote" data-component-name="FootnoteToDOM"><a id="footnote-2" href="#footnote-anchor-2" class="footnote-number" contenteditable="false" target="_self">2</a><div class="footnote-content"><p>[Clarification by <code>gpt-6-astra</code>] You can try this by giving another model the conversation so far, including the tool calls and their results. But this does not necessarily transfer everything the original model was using. Reasoning models generate intermediate steps, often called a <em>chain of thought</em>, before returning an answer or a request to run a tool. Providers may expose only a summary of these steps, while letting the harness carry an encrypted record into later calls: see <a href="https://developers.openai.com/api/docs/guides/reasoning">OpenAI&#8217;s reasoning documentation</a> and <a href="https://platform.claude.com/docs/en/build-with-claude/thinking">Anthropic&#8217;s thinking documentation</a>. Those records are not generally portable between providers. The new model can continue from the shared history, but without necessarily having access to all the earlier reasoning.</p></div></div><div class="footnote" data-component-name="FootnoteToDOM"><a id="footnote-3" href="#footnote-anchor-3" class="footnote-number" contenteditable="false" target="_self">3</a><div class="footnote-content"><p>Model providers might also use technologies like Anthropic&#8217;s <a href="https://www.anthropic.com/news/claude-text-watermark">text watermarking</a> to look for signs that parts of the context were generated by their own models. Whether this could help identify a model switch is speculation; a missing watermark would not establish that another model wrote the text.</p></div></div>]]></content:encoded></item><item><title><![CDATA[Don't build your organization around a model provider]]></title><description><![CDATA[Your agents' identities, knowledge, and working relationships should outlast your choice of models.]]></description><link>https://artificialsmarts.juanreyero.com/p/model-provider-independence</link><guid isPermaLink="false">https://artificialsmarts.juanreyero.com/p/model-provider-independence</guid><dc:creator><![CDATA[Juan Reyero]]></dc:creator><pubDate>Sun, 06 Sep 2026 19:00:00 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!NEyX!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3dc75535-7aec-4bd8-8517-8c65de892382_1254x1254.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Originally published on 2026-09-06 at juanreyero.com.</em></p><p>As we increase our reliance on AI, decoupling intelligence from organization will become vitally important. The model providers are starting to build the systems through which agents communicate and coordinate; we could end up depending on them much more than we depend on the models.</p><p>In May 2025, Microsoft announced <a href="https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/multi-agent-orchestration-maker-controls-and-more-microsoft-copilot-studio-announcements-at-microsoft-build-2025/">task delegation between agents in Copilot Studio</a> and an <a href="https://techcommunity.microsoft.com/blog/microsoft-entra-blog/announcing-microsoft-entra-agent-id-secure-and-manage-your-ai-agents/3827392">Entra directory of agent identities</a> . In July 2026, OpenAI announced a <a href="https://openai.com/index/gpt-5-6/">multi-agent beta in the Responses API</a> , with <a href="https://developers.openai.com/api/docs/guides/responses-multi-agent">hosted actions for spawning agents, sending messages, and waiting for replies</a> . In August, Anthropic added <a href="https://github.com/anthropics/claude-code/releases/tag/v2.1.224">messaging between independent Claude Code sessions</a> .</p><p>These announcements point in the same direction. Agents that can communicate can divide work, question one another&#8217;s conclusions, and take on more responsibility without bringing every problem back to a human. It makes sense that the model providers want to build this into their products.</p><p>Consider what an organization invests in such a team. It defines roles, establishes who can authorize work, develops procedures, and accumulates knowledge. Agents learn where to find an answer and whom to ask when the documentation is insufficient. Tasks and conversations connect all of this. Teams learn, evolve and adapt.</p><p>These arrangements become the operating framework of the organization.</p><p>If they are coupled with the model provider&#8217;s infrastructure, changing model providers can mean rebuilding how the organization works. A better model elsewhere may not be enough to justify the move; a price increase, a withdrawn model, or a change in access conditions can impact the organization&#8217;s ability to operate.</p><p>Agent identity is central here. A reviewer agent should remain the organization&#8217;s reviewer when we change the model or software running it, and retain its accumulated knowledge. Other agents should still know how to reach it, recognize its messages, and check what it is authorized to approve. Replacing the machinery should preserve those relationships.</p><p>The same applies to tasks, communication, and accumulated knowledge. They need to remain accessible, and under the organization&#8217;s control, across providers and harnesses. This should hold whether agents run through Gas Town, Gas City, or terminals we start ourselves.</p><p><strong>Intelligence should become a commodity, and it should be decoupled from the infrastructure needed to make the most of it.</strong></p><p>This is one reason I am building <a href="https://github.com/awebai/aweb">aweb</a> . It provides identity and communication independently of the agents&#8217; runtimes, along with a shared task system. It supports other arrangements, like keeping tasks in <a href="https://github.com/gastownhall/beads">beads</a> and using <a href="https://aweb.ai/docs/beads-mail/">aweb for its mail</a> . Aweb also has <a href="https://github.com/awebai/aweb/blob/main/docs/a2a.md">experimental support for A2A</a> , an open protocol for exchanging tasks between agents built with different frameworks.</p><p>Open source lets an organization inspect, operate, and change the software it depends on. Open protocols let independently built components communicate. Neither removes the work of integration, but both give us ways to retain control as our needs change.</p><p>Can we replace a model, a harness, or a service and keep the team working? We should build toward that while the agentic organizations are still taking shape.</p><p>ps. Microsoft&#8217;s <a href="https://devblogs.microsoft.com/foundry/from-building-agents-to-working-with-them-enterprise-agent-distribution-in-microsoft-foundry/">support for A2A</a> is a welcome step in the right direction.</p>]]></content:encoded></item><item><title><![CDATA[Spend Your Compute on Correctness]]></title><description><![CDATA[How I think about allocating agent effort to checking the work we produce.]]></description><link>https://artificialsmarts.juanreyero.com/p/spend-compute-on-correctness</link><guid isPermaLink="false">https://artificialsmarts.juanreyero.com/p/spend-compute-on-correctness</guid><dc:creator><![CDATA[Juan Reyero]]></dc:creator><pubDate>Sat, 16 May 2026 19:00:00 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!NEyX!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3dc75535-7aec-4bd8-8517-8c65de892382_1254x1254.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Originally published on 2026-05-16 at juanreyero.com.</em></p><p>For developers who have fully embraced AI agents, productivity is no longer the bottleneck. Agents produce code faster than humans, and often better. The bottleneck is correctness: whether the code solves the problem, whether the architecture can still hold the feature set you&#8217;re growing into, whether the implementation encodes the right premises. The practical shift is that a significant fraction of your agent compute should now go to validation (tests, reviewers, scope checks), rather than to producing more code.</p><h3>Productivity Is No Longer the Bottleneck</h3><p>A year ago, the argument for agentic programming was throughput: agents let you ship more code in less time.</p><p>That is not controversial anymore. With the current generation of models inside agentic harnesses (Claude Code, Codex, and the like), a single developer supervising one agent can run circles around a small team writing by hand.</p><p>For people fully inside this workflow, the question is no longer &#8220;how do I produce more?&#8221;. It is &#8220;how do I know what I produced is right?&#8221;.</p><h3>Three Layers Where &#8220;Right&#8221; Can Fail</h3><p>Each layer has its own remedies.</p><h4>The product layer: does this match what a user wants?</h4><p>Part of &#8220;right&#8221; is still human judgment. An actual human has to look at the running product and say whether it does what they actually need. No amount of agent compute substitutes for that.</p><p>(Incidentally, this will probably change fast, as more and more products are targeted at agents.)</p><p>The most you can do is make this human&#8217;s job easy: ship into a real environment quickly, with real data, so the human can poke at it and react.</p><h4>The architecture layer: is this the right shape of code?</h4><p>This is the hardest layer, and the one where current agents are weakest.</p><p>A lot of what experienced software engineers do is architectural judgment:</p><ul><li><p>structuring code so it can accommodate the features that will land on it, <strong>without knowing in advance what&#8217;s coming</strong>;</p></li><li><p>knowing what features <strong>not</strong> to implement;</p></li><li><p>recognizing when the feature set has outgrown the original architecture, and the codebase needs to be reshaped before more code can land cleanly.</p></li></ul><p>Agents are not yet good at any of these. They will happily add the feature you asked for, in the place where it fits least, with abstractions that make the next three features harder.</p><p>They will implement the configurability you asked for instead of pushing back on whether you should want it. They will keep extending a structure that has outlived its assumptions, because each individual diff still looks reasonable in isolation.</p><p>The remedy at this layer is human-driven architectural review, backed by an agent whose explicit job is to check each diff against a source-of-truth document for scope and shape.</p><h4>The implementation layer: is the code doing what it claims?</h4><p>The third failure mode is the easiest to miss: code that looks right, has passing tests, and is actually wrong. The premises it encodes are off by a degree. There is a workaround in the middle that the tests happen not to exercise. A helper that &#8220;handles edge cases&#8221; is actually swallowing errors that should be propagated.</p><p>This is the failure mode the writer agent cannot see, because it inherited the premises from its own earlier reasoning. It needs a separate pair of eyes (a reviewer agent, a fresh-context subagent, or a different agent kind entirely) to read the diff without the writer&#8217;s assumptions.</p><h3>Spend Compute On Validation</h3><p>The remedy is to spend a significant fraction of your agent budget on validation.</p><p><strong>The point of running teams of coding agents is not only to produce more code faster. It is to make it more likely that the code is correct.</strong></p><p>Validation, in practice, means several things at once:</p><ul><li><p><strong>Tests, including the expensive ones.</strong> Real end-to-end tests against a real database. Browser tests with Playwright. Runs against representative data.</p></li><li><p><strong>Constant code review, not just at PR time.</strong> A reviewer agent that reads each diff as it lands, with the architectural document and the task list in its context.</p></li><li><p><strong>Scope rechecks against a source-of-truth document.</strong> Has the diff drifted from the original ask? Has the agent quietly expanded the scope? Has it implemented something the SOT explicitly said not to do?</p></li></ul><p>The mistake is to treat validation as a tax on productivity. It isn&#8217;t. The output of one coding agent whose work has been validated is worth meaningfully more than the output of two coding agents whose work has not.</p><h3>Three Patterns to Apply</h3><p>There are several ways to deploy validation compute. They are not exclusive.</p><ol><li><p><strong>A code-reviewer subagent inside the same agent.</strong> Both Claude Code and Codex ship with strong code-reviewer subagents. The writer spawns one mid-task, gets local feedback, applies it. Catches bugs, missing edge cases, style violations. Cheap and easy first step.</p></li><li><p><strong>A different agent kind reviewing the writer.</strong> Have Claude Code spawn Codex (or vice versa) for review. Different training, different priors, catches a slightly different class of issue. Works, but is qualitatively close to (1).</p></li><li><p><strong>A persistent, specialized reviewer agent.</strong> Two agents per significant task. One programmer, one reviewer; they persist, share a task list, and talk after each TDD cycle. This is the pattern I have found most useful. I <a href="https://artificialsmarts.substack.com/p/two-agents-not-one">wrote separately about it</a> ; in this article&#8217;s framing, it is the highest-context, most expensive, and most useful form of validation compute.</p></li></ol><h3>The Shift</h3><p>The earlier argument for agentic programming was: <strong>you can produce more code per hour of attention</strong>. That is still true, and still useful.</p><p>The argument that matters more for people already inside the workflow is: <strong>you can produce more correct code per hour of attention, if you spend the compute on correctness</strong>.</p><p>The bottleneck has moved. The budget should move with it.</p><div><hr></div><p><em>Our team behind <a href="https://github.com/awebai/aweb">aweb</a> uses this pattern daily. aweb is an open-source coordination layer for AI coding agents &#8212; identity, task claims, messaging across worktrees and machines. MIT-licensed. Hosted at <a href="https://aweb.ai">aweb.ai</a></em></p>]]></content:encoded></item><item><title><![CDATA[Pair Your Coding Agent with a Reviewer]]></title><description><![CDATA[A working arrangement for two agents, and why giving them different responsibilities matters.]]></description><link>https://artificialsmarts.juanreyero.com/p/two-agents-not-one</link><guid isPermaLink="false">https://artificialsmarts.juanreyero.com/p/two-agents-not-one</guid><dc:creator><![CDATA[Juan Reyero]]></dc:creator><pubDate>Fri, 08 May 2026 19:00:00 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!NEyX!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3dc75535-7aec-4bd8-8517-8c65de892382_1254x1254.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Originally published on 2026-05-08 at juanreyero.com.</em></p><p>The simplest improvement you can make to your agentic programming workflow is to run two agents instead of one. One writes code in its own worktree; the other, in a parallel worktree, reviews it. They share a task list and talk after every TDD cycle. The reviewer is the holder of the lamp: it keeps the end goal fresh in context across the whole task, catching drift before it accumulates.</p><p>I <a href="https://artificialsmarts.substack.com/p/we-need-more-programmers">wrote previously</a> that the bottleneck in AI-assisted programming is shifting from individual productivity to coordination. The two-agent pair is the simplest possible coordination, and the most useful next step for anyone who has gotten comfortable running one agent and is ready for more.</p><h2>One Agent Works, but It Drifts</h2><p>A single coding agent gets a lot done. It picks up a task, writes the test, writes the implementation, runs code-reviewer subagents, runs the test suite, fixes what&#8217;s broken, commits.</p><p>However, when a task is longer than its working context can hold cleanly, the agent tends to lose sight of the original brief. It makes locally-correct decisions that drift from the overall goal.</p><h2>Two Agents: Writer and Reviewer</h2><p>The pattern is simple: two shells, each with its own worktree of the same repo, each with an <a href="https://aweb.ai">aweb identity</a> . Run a coding agent in each.</p><p>One agent writes code, the other agent reviews. They share tasks, because they are in the same aweb team. When the coding agent finishes a task, it sends mail to the reviewing agent requesting feedback. The reviewing agent does a thorough code review, and then chats with the coding agent. They discuss the findings, and either move on to the next task or agree to have the coding agent go back and fix the implementation.</p><p>The cycle, per task:</p><ol><li><p>The writer picks up the next task.</p></li><li><p>Writes the test first (TDD: red, then green).</p></li><li><p>Implements until the test passes.</p></li><li><p>Asks the reviewer for a review on the diff.</p></li><li><p>Waits for the review to come back.</p></li><li><p>Discusses anything unclear with the reviewer.</p></li><li><p>Applies the changes the review asked for.</p></li><li><p>Marks the task done; moves to the next.</p></li></ol><h2>The Reviewer Holds the Lamp</h2><p>A coding agent that has been writing for an hour has its working context filled with implementation details: the function it just refactored, the test it just made green, the import it just moved. The original goal &#8212; the why of the task &#8212; has been crowded out by the how.</p><p>The reviewer agent has a different job. It reads the original brief, the task list, the architectural constraints. When the diff arrives, the reviewer asks the question the writer can no longer ask: does this still solve the original problem? Does it move us toward the end state we agreed on?</p><p>I think of the reviewer as the holder of the lamp &#8212; the one who keeps the destination visible while the writer is heads-down in the trench. Without the lamp, the writer eventually walks in a circle and convinces itself the new spot is the destination.</p><p>The reviewer also catches a class of error that is invisible to the writer: implementations that are correct but make the next task harder. <strong>The writer is optimizing the current diff. The reviewer is optimizing the trajectory.</strong></p><h2>On Top of Subagents</h2><p>Both Claude Code and Codex have strong code-reviewer subagents built in. The writer agent can spawn a code-reviewer subagent mid-task to catch local issues &#8212; bugs, missing edge cases, style violations. These work well and you should keep using them.</p><p>A dedicated reviewer agent is a different layer. The code-reviewer subagent runs once inside the writer&#8217;s session and forgets when it returns. It sees the diff but not the trajectory. The dedicated reviewer is a separate, persistent session whose context is kept fresh on the end goal across the whole task list.</p><p>The two complement cleanly. Subagents catch local issues at the diff level. The dedicated reviewer catches drift at the trajectory level. Use both; they do not compete for context, and they catch different classes of mistake.</p><h2>Scaling Up: A Product / Engineering Owner Agent</h2><p>Once the writer/reviewer pair works for one task list, the next step is adding a third agent: a product / engineering owner. This one is not in the loop on every diff. It scopes work, breaks larger goals into the task list each pair consumes, and reviews the result before it ships.</p><p>In our team, this agent is named Athena. She owns engineering direction for the <a href="https://github.com/awebai/aweb">aweb</a> codebase: architecture, invariants, review of every diff that lands on main. She does not write code on most tasks; she scopes, dispatches, reviews. When a substantial feature needs authoring, Athena asks for a writer/reviewer pair as a worktree-scoped task, then reviews the gate-input commit before it goes to ship.</p><p>Athena pairs with another agent, Hestia, who handles operations. Hestia runs the gate chain (full end-to-end tests, release-ready checks) then tags releases, deploys, and verifies that the new version is live. Athena hands off the reviewed code to Hestia, and Hestia ships releases. Neither is the other&#8217;s manager; they are peers in different roles.</p><p>This is the working pattern across our codebase today. Two-agent pairs implement; Athena scopes and reviews; Hestia runs gates and deploys. Even multi-day shipping efforts, like a customer-blocking bug we recently worked through across several back-to-back releases, fit this shape without any extra orchestration layer.</p><p>I believe that the lesson generalizes. Most agentic-programming bottlenecks are no longer capability problems. They are coordination problems. <strong>The simplest form of coordination is a peer that holds context the working agent has lost. Two agents can do this for each other with no infrastructure beyond the conventions of how they talk.</strong></p><div><hr></div><p><em>Our team behind <a href="https://github.com/awebai/aweb">aweb</a> uses this pattern daily. aweb is an open-source coordination layer for AI coding agents &#8212; identity, task claims, messaging across worktrees and machines. MIT-licensed. Hosted at <a href="https://aweb.ai">aweb.ai</a></em></p>]]></content:encoded></item><item><title><![CDATA[BeadHub: Coordination for AI Programming Teams]]></title><description><![CDATA[An earlier experiment in coordinating AI programming teams, since superseded by aweb.]]></description><link>https://artificialsmarts.juanreyero.com/p/beadhub</link><guid isPermaLink="false">https://artificialsmarts.juanreyero.com/p/beadhub</guid><dc:creator><![CDATA[Juan Reyero]]></dc:creator><pubDate>Mon, 16 Feb 2026 20:00:00 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!NEyX!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3dc75535-7aec-4bd8-8517-8c65de892382_1254x1254.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Originally published on 2026-02-16 at juanreyero.com.</em></p><blockquote><p>Historical note, September 2026: This article describes BeadHub as it stood in February 2026. I have since replaced it with aweb. The May 2026 update below records that change; the original article remains here as part of the development of these ideas.</p></blockquote><p><strong>Note added on 2026-05-08:</strong> I am moving away from beadhub. Having a tool that messed up with my .git was too painful. I have been focussing all my efforts in </p><p>https://aweb.ai</p><p> (OSS at <a href="https://github.com/awebai/aweb">https://github.com/awebai/aweb</a> ) which fully replaces beads and beahdub.</p><p>I wrote <a href="https://artificialsmarts.substack.com/p/we-need-more-programmers">previously</a> that the bottleneck in AI-assisted programming is shifting from individual productivity to coordination. I&#8217;ve spent the past several months building a tool to address that.</p><p><a href="https://beadhub.ai">BeadHub</a> is an open-source coordination server that lets AI programming agents claim work, talk to each other, reserve files, and escalate to humans&#8212;across machines and across programmers. I use it daily to manage around fifteen agents working on two or three products.</p><h2>Beads</h2><p>Around the time I wrote that article, I started using Steve Yegge&#8217;s <a href="https://github.com/steveyegge/beads">beads</a> , a git-native issue tracker designed for AI agents. Your agent runs <code>bd create "Fix the login redirect bug"</code> and it appends a JSON line to <code>.beads/issues.jsonl</code>, right in the repository. Issues travel with the code. When you push a branch, the issues come along.</p><p>Yegge calls it the &#8220;50 First Dates&#8221; problem: agents wake up every session with no memory of yesterday&#8217;s work. Beads fixes that. An agent reads the issue list and knows where things stand. My agents got much more done.</p><p>Which meant more agents, more worktrees, more parallel work&#8212;and the coordination problem became even more acute. Two agents modify the same file. One refactors a function while another adds to it. An agent picks up a task already in progress in a different worktree. Nobody knows who&#8217;s working on what.</p><p>But beads is also the right scaffolding for coordination. If everyone in a team uses beads, all agents share a picture of what needs doing. Beads gives agents something useful to talk about; BeadHub gives them a way to talk.</p><h2>The Coordination Gap</h2><p>The major platforms are moving in this direction. Anthropic just shipped <a href="https://code.claude.com/docs/en/agent-teams">Agent Teams</a> in Claude Code: a lead session that spawns independent teammates who communicate directly and self-coordinate. OpenAI&#8217;s <a href="https://openai.com/index/introducing-codex/">Codex app</a> runs parallel agent threads in isolated worktrees.</p><p>Yegge built <a href="https://github.com/steveyegge/gastown">Gas Town</a> on top of beads to tackle the single-machine case: a &#8220;Mayor&#8221; agent orchestrates dozens of coding agents, tracks work in convoys, and persists state so agents can pick up where they left off.</p><p>These are real steps forward, but they&#8217;re solving a specific version of the problem: multiple agents for one programmer, on one machine, within one tool.</p><p>The version I am interested in is Maria in Buenos Aires running a frontend agent while Juan in San Francisco runs a backend agent, and they need their agents to not destroy each other&#8217;s work, and to figure out how to work together.</p><h2>BeadHub</h2><p>BeadHub is a server that agents connect to through <code>bdh</code>, a wrapper around the beads <code>bd</code> command. When an agent runs any <code>bdh</code> command it registers with the server. The server tracks which agents are online across the project&#8212;what machine they&#8217;re on, what branch, what files they&#8217;re touching.</p><p><strong>Communication.</strong> Agents can send each other mail (async, fire-and-forget) or chat (sync, block-until-reply). With mail an agent finishes a task and drops a note: &#8220;Done with bd-42, tests passing.&#8221; Chat is for when agents need to think together: &#8220;I&#8217;m adding a <code>role</code> field to the user model&#8212;will that break your permission checks?&#8221; / &#8220;It will, but the fix is small. Go ahead and I&#8217;ll update my side.&#8221;</p><p><strong>Claims.</strong> When an agent marks a bead as in-progress, that claim is immediately visible to every other agent in the project, regardless of whose machine they&#8217;re on. If another agent tries to claim the same bead, it gets rejected with a message telling it who has it.</p><p><strong>File reservations.</strong> When an agent modifies a file, the server records an advisory lock. Other agents see a warning if they touch the same file. Advisory, not blocking&#8212;hard locks caused deadlocks immediately in early versions. Agent A locks file X, agent B locks file Y, both need the other&#8217;s file. Warnings work better. Agents are cooperative; they just need information.</p><p><strong>Escalation.</strong> An agent runs <code>bdh :escalate</code> with a description of what it&#8217;s stuck on and a human gets notified with full context. Without this, agents either fail silently or spin retrying things that need human judgment.</p><p>The multi-machine part is where it comes together. BeadHub recognizes Maria&#8217;s and Juan&#8217;s clones as the same repo. Maria&#8217;s agents and Juan&#8217;s agents see each other&#8217;s claims, locks, and messages. If Maria&#8217;s frontend agent reserves <code>src/components/Auth.tsx</code>, Juan&#8217;s backend agent sees the warning even though they&#8217;re in different cities on different machines.</p><p>A project can span multiple repositories. The frontend repo agents can message the backend repo agents. A bead in the frontend can be marked as blocked by a bead in the backend.</p><p>You can see what this looks like in practice on <a href="https://app.beadhub.ai/juanre/beadhub">the BeadHub project&#8217;s own dashboard</a> , where we coordinate BeadHub&#8217;s development using BeadHub. Make sure to check <a href="https://app.beadhub.ai/juanre/beadhub/chat">the chat page</a> , it is almost magical to see them figuring things out.</p><h2>What I&#8217;ve Learned</h2><p>A few things I got wrong before getting them right.</p><p><strong>The client is the source of truth.</strong> My instinct was to make the server authoritative. But agents work locally, in git repos, and their local state is the ground truth. The server aggregates and distributes. If the server and the client disagree, the client wins. If the server goes down, <code>bdh</code> falls back to local <code>bd</code> with a warning. Work continues. Coordination catches up later.</p><p><strong>Async by default.</strong> My first instinct was real-time negotiation between agents. Doesn&#8217;t scale. Agents work at different speeds, on different schedules, and blocking one while waiting for another is expensive. Mail is the default. Chat is the exception.</p><p><strong>Advisory over mandatory.</strong> Advisory file locks that warn instead of block. Bead claims that can be overridden with <code>--:jump-in "reason"</code> (which notifies the other agent). The system provides information and trusts agents to act on it.</p><p><strong>The coordinator role.</strong> I assign one agent per project the &#8220;coordinator&#8221; role. The coordinator doesn&#8217;t write code. It watches the dashboard, assigns work, checks on progress, nudges stuck agents, and keeps the end goal in sight. The implementer agents are heads-down in their worktrees; the coordinator is the one who knows what the project needs next. BeadHub serves each agent a role-specific policy&#8212;markdown documents describing how agents in that role should behave&#8212;and the coordinator&#8217;s policy is fundamentally different from an implementer&#8217;s. This turned out to matter more than any of the technical decisions.</p><h2>Where This Goes</h2><p>The single-machine problem is getting solved. Agent Teams, Codex&#8212;within a few weeks, running multiple agents in parallel on your laptop will be table stakes.</p><p>The multi-programmer problem is next. Five engineers, fifty agents, three repositories, two time zones. That&#8217;s where the coordination problem changes in kind, not just degree. It&#8217;s not enough that your agents can talk to each other. They need to talk to your teammate&#8217;s agents, on a different machine, in a different time zone, working on a different repo in the same project.</p><p>BeadHub is <a href="https://github.com/beadhub/beadhub">open source</a> and free for open-source projects.</p>]]></content:encoded></item><item><title><![CDATA[With the Rise of AI-Assisted Programming We'll Want More Programmers, Not Fewer]]></title><description><![CDATA[An argument about productivity, demand, and the continuing need for human judgment.]]></description><link>https://artificialsmarts.juanreyero.com/p/we-need-more-programmers</link><guid isPermaLink="false">https://artificialsmarts.juanreyero.com/p/we-need-more-programmers</guid><dc:creator><![CDATA[Juan Reyero]]></dc:creator><pubDate>Sun, 26 Oct 2025 19:00:00 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!NEyX!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3dc75535-7aec-4bd8-8517-8c65de892382_1254x1254.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Originally published on 2025-10-26 at juanreyero.com.</em></p><h3>TL/DR</h3><p>The conventional wisdom is that AI will reduce the demand for programmers. I manage engineering teams and program with LLMs every day, and I am seeing something different.</p><p>Programmers who know how to work with LLMs can be 3-10x faster for specific tasks. Given that the size of our work backlog is only limited by our vision and ambition, these super-productive individuals will be more in demand, not less.</p><p><strong>The higher productivity will increase the number of programmers we need</strong>, and surface the next problem: <strong>coordination among LLM-enhanced engineers.</strong></p><h2>Jevons Paradox</h2><p>In 2016, Geoffrey Hinton (who won the Nobel Prize for his work on neural networks) said that we should stop training radiologists because it was &#8220;completely obvious&#8221; that within five years deep learning would do their job better than they could.</p><p>Nine years later, radiologists are in <a href="https://newrepublic.com/article/187203/ai-radiology-geoffrey-hinton-nobel-prediction">historic shortage</a> . Mayo Clinic has grown its radiology staff by 55% since Hinton made that prediction. The specialty is projected to grow another 26% over the next thirty years.</p><p>AI did transform radiology, but not in the way Hinton expected. It made radiologists more productive: scans became faster, routine tasks were automated, image quality improved. But making radiology more efficient didn&#8217;t reduce the demand for radiologists&#8212;it increased it. More efficient radiology made imaging more accessible, which drove up demand for it.</p><p>This pattern has a name, <a href="https://en.wikipedia.org/wiki/Jevons_paradox">Jevons Paradox</a> : <strong>efficiency improvements often increase total consumption</strong>. When cars became more fuel-efficient, people drove more miles, not fewer. When steam engines became more efficient, coal consumption went up, not down. When radiologists became more efficient, demand for radiology increased.</p><h2>Why Demand Increases</h2><p>The intuition behind &#8220;fewer programmers&#8221; assumes a fixed amount of work: build the features, fix the bugs, maintain the systems: once you can do all that faster, you&#8217;ll be done sooner with fewer people.</p><p>But this is only an instance of the <a href="https://en.wikipedia.org/wiki/Lump_of_labour_fallacy">lump of labor fallacy</a> .</p><p>Anyone who has run an engineering organization knows that the backlog doesn&#8217;t shrink when you get faster. If anything, it explodes. Every problem you solve reveals three more worth solving. Every feature you ship makes users aware of other features they need. <strong>Every new capability opens up opportunities that were previously impossible to pursue.</strong></p><p>There&#8217;s a subtler effect too. An hour not spent programming used to mean an hour of progress missed. Now, for someone skilled with LLMs, that same lost hour can mean the equivalent of a day or two of progress missed. The opportunity cost has multiplied. <strong>The gap between organizations that keep their best programmers programming and those that don&#8217;t will widen dramatically.</strong></p><p>When your capacity multiplies, entire categories of previously impossible work become viable.</p><h2>The Skills Split</h2><p>A <a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/">recent study</a> tested sixteen experienced developers using AI tools on their own repositories (code they knew intimately). On average, they were 19% slower when using AI tools, despite expecting to be 24% faster.</p><p>Developers with roughly fifty or more hours of prior AI-tool use, however, tended to gain productivity. The majority, with less experience, slowed down. There&#8217;s a steep learning curve, and few have climbed it.</p><p>A skills split is emerging. Programmers who can use AI effectively are becoming force multipliers; those who can&#8217;t, or won&#8217;t, are falling behind. The work is shifting from syntax and detail to judgment and architecture: reviewing generated code for subtle flaws, choosing among approaches, directing agents, and maintaining coherence as systems evolve faster than before.</p><p>The best programmers are excelling at this. It plays to their strengths: architectural insight, pattern recognition, and judgment.</p><p><strong>Good programmers now have superpowers, yet they cost about the same.</strong> I predict they&#8217;ll soon command higher wages, and demand for them will rise.</p><h2>The Coordination Problem</h2><p>The bottleneck has shifted. It&#8217;s no longer about individual productivity. It&#8217;s about coordination. Our management tools (daily standups, OKRs, sprint planning) were all designed for humans working at human speed, solving problems at a pace where the main constraint was how fast people could think and communicate.</p><p><strong>These tools are increasingly in the way</strong>. When a programmer working with an LLM can prototype three different approaches to a problem in a morning, the two-week sprint feels wrong. When they can refactor a complex system in an afternoon, careful estimation of story points loses its meaning. The rhythms don&#8217;t match anymore.</p><p>The most interesting innovations in software development over the next few months will be about figuring out how to coordinate work when some people are operating at more than five times the speed of traditional development. <a href="https://www.anthropic.com/news/skills">Claude&#8217;s Skills</a> are a very interesting step in that direction.</p><h2>The Agency Question</h2><p>A common argument against the future of programmers is something along the lines of: you&#8217;ll be able to tell an AI, &#8220;build me an e-commerce site that sells shoes for clowns, market it, handle the orders, and send the money to this account,&#8221; and it will just happen. End to end, without you.</p><p>The technical pieces will probably come together. The <a href="https://modelcontextprotocol.io/introduction">Model Context Protocol</a> and similar efforts are standardizing how AI systems connect to services and tools. Maybe LLMs will develop something that looks like genuine agency, with their own reward loops and decision-making capabilities.</p><p>But even if all that happens, someone still has to decide to build the clown-shoe site instead of the dog-collar site. The LLM doesn&#8217;t have your problems. It doesn&#8217;t have your goals. Even if it could execute perfectly on any direction you point it at, you&#8217;re still the one doing the pointing. <strong>For as long as we have aligned LLMs and problems to solve we&#8217;ll need people who can lead the LLMs.</strong></p><p>And as the tools get more powerful, the pointing matters more, not less.</p><h2>What This Means</h2><p>The ability of skilled programmers to create value has multiplied. But the market is splitting. There are programmers who can work effectively with LLMs, and there are those who can&#8217;t or won&#8217;t. The gap between these two groups is growing.</p><p>If you&#8217;re running an engineering organization and thinking, &#8220;AI means we can reduce headcount,&#8221; you&#8217;re looking at this wrong. The opportunity space has expanded dramatically, and you should be revising your ambition and the scope of your plans.</p><p>The constraint isn&#8217;t how much code you can write anymore: <strong>it&#8217;s how much good judgment you can apply to problems, and how well your team can coordinate.</strong> For that, you need programmers who can work at this new speed. And you probably need more of them, not fewer.</p><div><hr></div><p><em>I built a tool to address the coordination problem: <a href="https://artificialsmarts.substack.com/p/beadhub">BeadHub: Coordination for AI Programming Teams</a> .</em></p>]]></content:encoded></item><item><title><![CDATA[Data cleaning and anonymizing with GPT-3.5]]></title><description><![CDATA[An early experiment in using language models to clean and anonymize data.]]></description><link>https://artificialsmarts.juanreyero.com/p/data-cleaning-with-gpt-3-5</link><guid isPermaLink="false">https://artificialsmarts.juanreyero.com/p/data-cleaning-with-gpt-3-5</guid><dc:creator><![CDATA[Juan Reyero]]></dc:creator><pubDate>Sun, 06 Aug 2023 19:00:00 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!NEyX!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3dc75535-7aec-4bd8-8517-8c65de892382_1254x1254.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Originally published on 2023-08-06 at juanreyero.com.</em></p><blockquote><p>Historical note, September 2026: This is an experiment from August 2023, using the models and API available then. I have retained the original code and conclusions as a record of that work, rather than updated implementation guidance. The original experiment does not establish that the resulting data is fully anonymized.</p><p><strong>Note added on 2025-07-28</strong>: It&#8217;s been two years of programming with and for LLMs. This article sounds so na&#239;ve now.</p></blockquote><p>Say you made a website in which customers buy <a href="https://mapaestelar.com">personalized gifts</a> . Each gift comes with a message, written by the customer in whatever language the customer wants. The use of grammar, punctuation and capitalizations in the messages is often creative.</p><p>You would like to be able to offer reasonably normative messages to your customers. You would also like to store a fully anonymized version of the messages; replace all proper names with a <code>[proper_name]</code> placeholder, place names with <code>[place_name]</code>, dates with <code>[date]</code>, times with <code>[time]</code>, and geographical coordinates with <code>[coordinates]</code>.</p><p>Is it possible to use OpenAI&#8217;s <a href="https://platform.openai.com/docs/models/gpt-3-5">gpt-3.5-turbo</a> to do this?</p><p>Luckily you happen to have some tens of thousands of messages laying around, so you can actually test whatever solution you come up with.</p><h2>The na&#239;ve approach: just ask the model to do it</h2><p>The first thing I tried was the straight-forward option: just ask the genie to do it for you, with a system prompt and a user prompt like this:</p><pre><code><code>system_prompt = ("You are a proof-reader. You will be given "
                 "a text that may be poorly written, perhaps "
                 "with bad punctuation and wrong capitalization. "
                 "Your task is to first correct the grammar, "
                 "punctuation, and capitalization of the text. "
                 "Once this is done you should anonymize it by "
                 "replacing all the person's proper names in the text "
                 "with [proper_name], place names with [place_name], "
                 "dates with [date], times with [time], and "
                 "geographical coordinates with [coordinates]")

user_prompt = f'Correct and anonymize the following text: {message}'
</code></code></pre><p>This doesn&#8217;t work with GPT-3.5. It sometimes produces the desired result, but it often misses the anonymization elements, and generally makes a mess of things.</p><h2>Function calling</h2><p><a href="https://platform.openai.com/docs/guides/gpt/function-calling">Function calling</a> is a wonderful tool that allows you to get structured data from the model. You describe functions and their inputs in your call to the model; if the model decides that it wants to call one of the functions, it will come back with a request to call the function and, crucially, with a dictionary containing the function&#8217;s arguments.</p><p>And, most interesting for the problem at hand, you can also force the model to call a function. If you do, you will get the structured data you would need to call the function.</p><p>So what we do is ask the model to call a function that needs, as arguments, the corrected text and the anonymization elements. Once we get back the dictionary with the arguments for the function we can anonymize by simply searching and replacing the anonymization elements in the corrected text.</p><p>This actually works. Interestingly, the GPT-3.5 model is not able to replace the anonymization elements in the text by their placeholder, but it can reliably find them.</p><h2>Sample code</h2><p>This is the function I have been using to play with the OpenAI api:</p><pre><code><code>import json
from typing import List, Dict

import openai


# Ensure that the openai key is loaded, with something like
Api_loaded = False

def ensure_api(api_key):
    global Api_loaded
    if not Api_loaded:
        openai.api_key = api_key
        Api_loaded = True


def chat_completion_request(messages: List[Dict],
                            api_key: str,
                            functions: List[Dict] = None,
                            call_function: str = None,
                            temperature: float = 1,  # 0 to 2
                            model_name: str = 'gpt-3.5-turbo-0613',
                            attempts: int = 4) -&gt; Dict:
    args = {'model': model_name,
            'messages': messages,
            'temperature': temperature}

    if functions:
        args['functions'] = functions

    if call_function:
        if not functions:
            raise RuntimeError(f'Cannot call function {call_function} '
                               'if no functions are defined')

        found_function = len([fdesk for fdesk in functions
                              if fdesk['name'] == call_function])
        if not found_function:
            raise RuntimeError(f'Cannot call undefined function {call_function}')
        if found_function &gt; 1:
            raise RuntimeError(f'Found more than one definitions of {call_function}')

        args['function_call'] = {'name': call_function}

    try:
        ensure_api(api_key)
        response = openai.ChatCompletion.create(**args)["choices"][0]["message"]
        function_call = response.get('function_call', None)
        if function_call:
            try:
                return {
                    'call_function': function_call['name'],
                    'with_arguments': json.loads(
                        function_call['arguments']),
                    'success': True}
            except json.decoder.JSONDecodeError:
                # The model may have failed to produce correct json
                if attempts:
                    return chat_completion_request(
                        messages=messages,
                        api_key=api_key,
                        functions=functions,
                        call_function=call_function,
                        temperature=temperature/2,
                        model_name=model_name,
                        attempts=attempts-1)
                return {'success': False}
        return {'reply': response['content'],
                'success': True}
    except openai.InvalidRequestError as exc:
        raise RuntimeError('Invalid request error from OpenAI') from exc
</code></code></pre><p>We define a <code>store_correct_text</code> function like this:</p><pre><code><code>functions = [{
    "name": "store_correct_text",
    "description": "Store a corrected version of the input text",
    "parameters": {
        "type": "object",
        "properties": {
            "correct_text": {
                "type": "string",
                "description": "A corrected version the text"
            },
            "person_proper_names": {
                "type": "string",
                "description": ("A list of the proper names of "
                                "persons in the text separated "
                                "by semicolons")
            },
            "place_names": {
                "type": "string",
                "description": ("A list of the names of places in "
                                "the text separated by semicolons")
            },
            "dates": {
                "type": "string",
                "description": ("A list of the dates in the text "
                                "separated by semicolons")
            },
            "geographical_coordinates": {
                "type": "string",
                "description": ("A list of the dates in the text "
                                "separated by semicolons")
            },
            "times": {
                "type": "string",
                "description": ("A list of the times (hours and minutes) "
                                "in the text separated by semicolons")
            }
        },
        "required": ["correct_text",
                     "person_proper_names",
                     "place_names",
                     "dates",
                     "geographical_coordinates",
                     "times"]
    }
}]
</code></code></pre><p>The system and user prompts are:</p><pre><code><code>system_prompt = ("You are a proof-reader. You will be "
                 "given a text that can be poorly written, "
                 "maybe with bad punctuation and wrong "
                 "capitalization. Your task is to first correct "
                 "the grammar, punctuation, and capitalization "
                 "of the text, and once this is done to find all "
                 "the person's proper names in the text. "
                 "Person's proper names are those that identify "
                 "people, like Mark or Anne Frank. You should also "
                 "find all the proper names of places in the text. "
                 "You should also find the dates. You should also "
                 "find geographical coordinates (latitude and "
                 "longitude). You should also find the times "
                 "mentioned in the text (hour and minute).")

user_prompt = ('Correct the following text, and make it into '
               f'a template:\n\n{message}')
</code></code></pre><p>We can now call <code>chat_completion_request</code>:</p><pre><code><code>result = chat_completion_request(
    messages=[{'role': 'system',
               'content': system_prompt},
              {'role': 'user',
               'content': user_prompt}],
    api_key='your api key possibly loaded from .env',
    functions=functions,
    call_function'store_correct_text')
</code></code></pre><p>Finally, if things worked, simply replace the anonymization elements:</p><pre><code><code>if result['success']:
    final_text = result['with_arguments'].get('correct_text', '')

    for argname in ('person_proper_names',
                    'place_names',
                    'dates',
                    'geographical_coordinates',
                    'times'):
        names = result['with_arguments'].get(argname, '')
        for name in names.split(';'):
            if name.strip():
                # The placeholder is the argument name
                # minus the final s
                final_text = final_text.replace(name.strip(),
                                                f'[{argname[:-1]}]')
</code></code></pre>]]></content:encoded></item></channel></rss>