<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>llm=claude-sonnet-4.6 on tomrochette.com</title>
    <link>https://tomrochette.com/tags/llm=claude-sonnet-4.6/</link>
    <description>Recent content in llm=claude-sonnet-4.6 on tomrochette.com</description>
    <generator>Hugo -- gohugo.io</generator>
    <language>en</language>
    <managingEditor>tom@tomrochette.com (Tom Rochette)</managingEditor>
    <webMaster>tom@tomrochette.com (Tom Rochette)</webMaster>
    <copyright>© 2026 Tom Rochette</copyright>
    <lastBuildDate>Fri, 02 Oct 2026 19:42:41 -0400</lastBuildDate><atom:link href="https://tomrochette.com/tags/llm=claude-sonnet-4.6/index.xml" rel="self" type="application/rss+xml" />
    
    <item>
      <title>What Changes When Coding Agents Are Infinitely Parallel</title>
      <link>https://tomrochette.com/agi/llm/what-changes-when-coding-agents-are-infinitely-parallel/</link>
      <pubDate>Wed, 18 Feb 2026 00:00:00 +0000</pubDate>
      <author>tom@tomrochette.com (Tom Rochette)</author>
      <guid>https://tomrochette.com/agi/llm/what-changes-when-coding-agents-are-infinitely-parallel/</guid>
      <category>artificial-general-intelligence</category><category>fully-ai-generated</category><category>llm=claude-sonnet-4.6</category><category>llm=glm-5.3</category><category>llm=deepseek-v4.1-flash</category>
      <description>&lt;p&gt;Imagine you can run hundreds or thousands of coding agents in parallel.&#xA;How would you use them?&#xA;&lt;strong&gt;The interesting answer is not &amp;ldquo;do everything faster&amp;rdquo;, but rather what fundamentally changes when parallelism is cheap.&lt;/strong&gt;&lt;/p&gt;&#xA;&#xA;&lt;h2 class=&#34;relative group&#34;&gt;From Sequential Exploration to Parallel Search&#xA;    &lt;div id=&#34;from-sequential-exploration-to-parallel-search&#34; class=&#34;anchor&#34;&gt;&lt;/div&gt;&#xA;    &#xA;    &lt;span&#xA;        class=&#34;absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none&#34;&gt;&#xA;        &lt;a class=&#34;text-primary-300 dark:text-neutral-700 !no-underline&#34; href=&#34;#from-sequential-exploration-to-parallel-search&#34; aria-label=&#34;Anchor&#34;&gt;#&lt;/a&gt;&#xA;    &lt;/span&gt;&#xA;    &#xA;&lt;/h2&gt;&#xA;&lt;p&gt;When you have one agent, you think carefully before acting, because going down dead ends is expensive.&#xA;With thousands of agents, the trade-off reverses.&#xA;&lt;strong&gt;Exploration becomes cheap, and convergence becomes the hard problem.&lt;/strong&gt;&#xA;Your job shifts from &amp;ldquo;what should I try?&amp;rdquo; to &amp;ldquo;how do I synthesize thousands of results?&amp;rdquo;&lt;/p&gt;&#xA;&#xA;&lt;h2 class=&#34;relative group&#34;&gt;High-Leverage Use Patterns&#xA;    &lt;div id=&#34;high-leverage-use-patterns&#34; class=&#34;anchor&#34;&gt;&lt;/div&gt;&#xA;    &#xA;    &lt;span&#xA;        class=&#34;absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none&#34;&gt;&#xA;        &lt;a class=&#34;text-primary-300 dark:text-neutral-700 !no-underline&#34; href=&#34;#high-leverage-use-patterns&#34; aria-label=&#34;Anchor&#34;&gt;#&lt;/a&gt;&#xA;    &lt;/span&gt;&#xA;    &#xA;&lt;/h2&gt;&#xA;&lt;p&gt;&lt;strong&gt;Speculative execution on decisions.&lt;/strong&gt;&#xA;At every architectural fork, such as &amp;ldquo;should this be a queue-based system or polling?&amp;rdquo;, you do not decide, you branch.&#xA;Two fleets of agents build both options, and you evaluate the results, borrowing the processor idea of &lt;a href=&#34;https://en.wikipedia.org/wiki/Speculative_execution&#34;  target=&#34;_blank&#34; rel=&#34;noreferrer&#34;&gt;&lt;img class=&#34;external-link-favicon&#34; src=&#34;https://www.google.com/s2/favicons?domain=en.wikipedia.org&amp;sz=128&#34; alt=&#34;&#34; width=&#34;16&#34; height=&#34;16&#34; loading=&#34;lazy&#34;&gt;speculative execution&lt;/a&gt;.&#xA;This pattern is very useful for situations where you genuinely do not know which approach is better until you have tried it.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Mutation testing at scale.&lt;/strong&gt;&#xA;Spin up hundreds of agents making targeted changes to a codebase, each trying a different hypothesis about why a bug exists or how to optimize something.&#xA;It is like &lt;a href=&#34;https://en.wikipedia.org/wiki/Fuzzing&#34;  target=&#34;_blank&#34; rel=&#34;noreferrer&#34;&gt;&lt;img class=&#34;external-link-favicon&#34; src=&#34;https://www.google.com/s2/favicons?domain=en.wikipedia.org&amp;sz=128&#34; alt=&#34;&#34; width=&#34;16&#34; height=&#34;16&#34; loading=&#34;lazy&#34;&gt;fuzzing&lt;/a&gt;, but semantically directed: instead of random inputs, each change probes a specific theory.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Full-stack consistency checking.&lt;/strong&gt;&#xA;Have agents simultaneously maintain the contract between every pair of services in your system, constantly verifying that implementations match specs, that error handling is symmetric, and that naming is consistent.&#xA;Sequential review often misses these checks.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Competitive benchmarking of approaches.&lt;/strong&gt;&#xA;For a problem like network latency or caching strategy, you could have 50 agents implement 50 different approaches against the same test harness, and just pick the winner.&#xA;No reading papers and reasoning about tradeoffs, you settle the question empirically.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Living documentation.&lt;/strong&gt;&#xA;Agents continuously reconcile docs, comments, and code.&#xA;Every pull request triggers agents that check for documentation drift, update runbooks, and surface inconsistencies.&lt;/p&gt;&#xA;&#xA;&lt;h2 class=&#34;relative group&#34;&gt;The Harder Problems Cheap Parallelism Creates&#xA;    &lt;div id=&#34;the-harder-problems-cheap-parallelism-creates&#34; class=&#34;anchor&#34;&gt;&lt;/div&gt;&#xA;    &#xA;    &lt;span&#xA;        class=&#34;absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none&#34;&gt;&#xA;        &lt;a class=&#34;text-primary-300 dark:text-neutral-700 !no-underline&#34; href=&#34;#the-harder-problems-cheap-parallelism-creates&#34; aria-label=&#34;Anchor&#34;&gt;#&lt;/a&gt;&#xA;    &lt;/span&gt;&#xA;    &#xA;&lt;/h2&gt;&#xA;&lt;p&gt;&lt;strong&gt;Synthesis is the bottleneck.&lt;/strong&gt;&#xA;If 1,000 agents each produce a pull request, you are back to being the serial bottleneck reviewing them.&#xA;You need meta-agents whose job is to evaluate and rank the output of other agents, with clear scoring functions.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;State and conflict.&lt;/strong&gt;&#xA;Agents working in parallel on the same codebase will conflict.&#xA;You probably want agents working in isolated sandboxes (branches, ephemeral clusters, test environments) with a merge or tournament layer on top.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Task decomposition quality matters more, not less.&lt;/strong&gt;&#xA;Bad task specs mean 1,000 agents going confidently in the wrong direction simultaneously.&#xA;The skill of writing tight, evaluable task specs becomes enormously valuable.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Evaluation functions become critical.&lt;/strong&gt;&#xA;&amp;ldquo;Did the agent succeed?&amp;rdquo; needs a concrete, automated answer.&#xA;You cannot manually review 1,000 outputs.&#xA;That constraint pushes you toward &lt;a href=&#34;https://en.wikipedia.org/wiki/Test-driven_development&#34;  target=&#34;_blank&#34; rel=&#34;noreferrer&#34;&gt;&lt;img class=&#34;external-link-favicon&#34; src=&#34;https://www.google.com/s2/favicons?domain=en.wikipedia.org&amp;sz=128&#34; alt=&#34;&#34; width=&#34;16&#34; height=&#34;16&#34; loading=&#34;lazy&#34;&gt;test-driven development&lt;/a&gt;.&lt;/p&gt;&#xA;&#xA;&lt;h2 class=&#34;relative group&#34;&gt;What This Implies For Infrastructure Work&#xA;    &lt;div id=&#34;what-this-implies-for-infrastructure-work&#34; class=&#34;anchor&#34;&gt;&lt;/div&gt;&#xA;    &#xA;    &lt;span&#xA;        class=&#34;absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none&#34;&gt;&#xA;        &lt;a class=&#34;text-primary-300 dark:text-neutral-700 !no-underline&#34; href=&#34;#what-this-implies-for-infrastructure-work&#34; aria-label=&#34;Anchor&#34;&gt;#&lt;/a&gt;&#xA;    &lt;/span&gt;&#xA;    &#xA;&lt;/h2&gt;&#xA;&lt;p&gt;The most compelling application is &lt;strong&gt;parallel experimentation on actual infrastructure parameters&lt;/strong&gt;.&#xA;Spin up ephemeral environments, have agents try different configurations or scheduling strategies, measure actual performance, then tear the environments down.&#xA;What makes infrastructure hard is that you normally cannot afford to run 50 experiments simultaneously.&#xA;With cheap parallel agents managing the scaffolding, that constraint disappears.&lt;/p&gt;&#xA;&lt;p&gt;The practical limit ends up being &lt;strong&gt;compute and money, not ideas&lt;/strong&gt;, which is a fundamentally different world than the one most engineering workflows were designed for.&lt;/p&gt;&#xA;</description>
      
    </item>
    
  </channel>
</rss>
