Skip to content

an open-source alternative to the dots, bots and muses of the world, run an office of claude code/codex like agents on your laptop, sandboxes or anywhere, uses your existing subscriptions

32.3 KB
Raw Blame
<!doctype html>
<html lang="en">
<head>
<!-- Google tag (gtag.js), Google Ads AW-18388216863. Loaded only when Do Not Track is
     off, so it follows the same rule as the PostHog block on this site. -->
<script>
  window.dataLayer = window.dataLayer || [];
  function gtag(){ dataLayer.push(arguments); }
  if (navigator.doNotTrack !== '1') {
    var gtagScript = document.createElement('script');
    gtagScript.async = true;
    gtagScript.src = 'https://www.googletagmanager.com/gtag/js?id=AW-18388216863';
    document.head.appendChild(gtagScript);
    gtag('js', new Date());
    gtag('config', 'AW-18388216863');
  }
</script>

<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>Context Engineering for AI Agents: Budgeting the Window — Munder Difflin Blog</title>
<meta name="description" content="Why context engineering beats prompt wording for agents — and the tactics (isolation, retrieval, externalized state, compaction) that keep the window lean." />
<link rel="canonical" href="https://munderdiffl.in/blog/context-engineering-for-ai-agents/" />

<link rel="icon" type="image/png" href="https://munderdiffl.in/logo.png" />
<link rel="apple-touch-icon" href="https://munderdiffl.in/logo.png" />
<meta name="theme-color" content="#F5F2E8" />

<!-- Open Graph -->
<meta property="og:site_name" content="Munder Difflin Blog" />
<meta property="og:type" content="article" />
<meta property="og:title" content="Context Engineering for AI Agents: Budgeting the Window — Munder Difflin Blog" />
<meta property="og:description" content="Why context engineering beats prompt wording for agents — and the tactics (isolation, retrieval, externalized state, compaction) that keep the window lean." />
<meta property="og:url" content="https://munderdiffl.in/blog/context-engineering-for-ai-agents/" />
<meta property="og:image" content="https://munderdiffl.in/blog/assets/media/context-engineering-for-ai-agents/hero.png" />
<meta property="og:image:width" content="1600" />
<meta property="og:image:height" content="900" />
<meta property="og:image:alt" content="Illustration: Context Engineering for AI Agents: Budgeting the Window" />
<meta property="og:locale" content="en_US" />
<meta property="article:published_time" content="2026-06-04T00:00:00.000Z" />
<meta property="article:author" content="Chaitanya Giri" />
<meta property="article:tag" content="Context Engineering" />
<meta property="article:tag" content="Memory" />
<meta property="article:tag" content="Multi-Agent" />
<meta property="article:tag" content="Claude Code" />


<!-- Twitter -->
<meta name="twitter:card" content="summary_large_image" />
<meta name="twitter:title" content="Context Engineering for AI Agents: Budgeting the Window — Munder Difflin Blog" />
<meta name="twitter:description" content="Why context engineering beats prompt wording for agents — and the tactics (isolation, retrieval, externalized state, compaction) that keep the window lean." />
<meta name="twitter:image" content="https://munderdiffl.in/blog/assets/media/context-engineering-for-ai-agents/hero.png" />
<meta name="twitter:image:alt" content="Illustration: Context Engineering for AI Agents: Budgeting the Window" />

<link rel="alternate" type="application/atom+xml" title="Munder Difflin Blog" href="https://munderdiffl.in/blog/feed.xml" />
<link rel="preconnect" href="https://fonts.googleapis.com" />
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
<link href="https://fonts.googleapis.com/css2?family=Fraunces:opsz,[email protected],500;9..144,600;9..144,700;9..144,900&amp;family=Source+Serif+4:opsz,[email protected],400;8..60,600;8..60,700&amp;family=Space+Grotesk:wght@500;600;700&amp;family=JetBrains+Mono:wght@400;600&amp;display=swap" rel="stylesheet" />
<link rel="stylesheet" href="/blog/assets/blog-press.css" />

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "BlogPosting",
  "headline": "Context Engineering for AI Agents: Budgeting the Window",
  "description": "Why context engineering beats prompt wording for agents — and the tactics (isolation, retrieval, externalized state, compaction) that keep the window lean.",
  "image": ["https://munderdiffl.in/blog/assets/media/context-engineering-for-ai-agents/hero.png"],
  "datePublished": "2026-06-04T00:00:00.000Z",
  "dateModified": "2026-06-04T00:00:00.000Z",
  "author": { "@type": "Person", "name": "Chaitanya Giri" },
  "publisher": {
    "@type": "Organization",
    "name": "Munder Difflin",
    "logo": { "@type": "ImageObject", "url": "https://munderdiffl.in/logo.png" }
  },
  "mainEntityOfPage": { "@type": "WebPage", "@id": "https://munderdiffl.in/blog/context-engineering-for-ai-agents/" },
  "keywords": "context engineering, context engineering for ai agents, context window management, agent context window, context rot",
  "articleSection": "Guides",
  "inLanguage": "en-US",
  "url": "https://munderdiffl.in/blog/context-engineering-for-ai-agents/"
}
</script>
<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "BreadcrumbList",
  "itemListElement": [
    { "@type": "ListItem", "position": 1, "name": "Home", "item": "https://munderdiffl.in/" },
    { "@type": "ListItem", "position": 2, "name": "Blog", "item": "https://munderdiffl.in/blog/" },
    { "@type": "ListItem", "position": 3, "name": "Context Engineering for AI Agents: Budgeting the Window", "item": "https://munderdiffl.in/blog/context-engineering-for-ai-agents/" }
  ]
}
</script>

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    { "@type": "Question", "name": "What is context engineering?", "acceptedAnswer": { "@type": "Answer", "text": "Context engineering is the discipline of deciding exactly what tokens occupy a model's finite context window at each step — instructions, tools, retrieved memory, and history — so the model has what it needs and nothing it doesn't. It's the system-level successor to prompt engineering: prompt engineering tunes the wording of one message, context engineering curates the whole window over a long-running, multi-turn agent." } },
    { "@type": "Question", "name": "How is context engineering different from prompt engineering?", "acceptedAnswer": { "@type": "Answer", "text": "Prompt engineering optimizes a single static prompt. Context engineering manages a dynamic, growing window across many turns and often many agents — what to load, what to retrieve just-in-time, what to summarize, and what to keep out of the window entirely. For agents that run for hours and call dozens of tools, the prompt is the smallest part of the problem." } },
    { "@type": "Question", "name": "What is context rot?", "acceptedAnswer": { "@type": "Answer", "text": "Context rot is the degradation in answer quality as a context window fills up. More tokens isn't strictly better: irrelevant history, stale tool output, and contradictory notes distract the model and crowd out what matters. A large window is a budget to spend carefully, not a bucket to fill." } },
    { "@type": "Question", "name": "Why does context engineering matter more in a multi-agent system?", "acceptedAnswer": { "@type": "Answer", "text": "Because every agent has its own finite window and they must coordinate. If you pour all shared state into every agent's context, you pay for it in tokens, latency, and confusion. The fix is to give each agent a small, role-scoped window and keep the shared state — mailboxes, logs, the plan — in files the agent reads on demand." } }
    
  ]
}
</script>


<script>
  // Anonymous blog analytics (PostHog) — same posture as the main site and the
  // app's TELEMETRY.md: public write-only token ('' = inert), anonymous-only
  // (person_profiles 'never'), no autocapture/recording, DNT respected.
  (function () {
    var PH_KEY = 'phc_qJzsokwcbtTGHemnvywR3nSqivNNgMjHeNVEWkZNydFB';
    if (!PH_KEY || navigator.doNotTrack === '1') return;
    var s = document.createElement('script');
    s.async = true;
    s.src = 'https://us-assets.i.posthog.com/static/array.js';
    s.onload = function () {
      if (!window.posthog) return;
      window.posthog.init(PH_KEY, {
        api_host: 'https://us.i.posthog.com',
        person_profiles: 'never',
        autocapture: false,
        before_send: function (event) {
          // PostHog fills $ip from the connection unless the event carries one.
          // Null it here and disable the GeoIP lookup so no IP, city, postal
          // code or lat/long is ever derived or stored for this page view.
          if (event && event.properties) {
            event.properties.$ip = null;
            event.properties.$geoip_disable = true;
          }
          return event;
        },
        disable_session_recording: true,
        respect_dnt: true
      });
    };
    document.head.appendChild(s);
  })();
</script>
</head>
<body class="theme-press">
<a class="skip-link" href="#main">Skip to content</a>
<nav id="nav">
  <a class="brand" href="https://munderdiffl.in" aria-label="Munder Difflin home">
    <img class="brand-logo" src="https://munderdiffl.in/logo.png" width="1318" height="840" alt="Munder Difflin, Inc. — Multi-Agent Harness" />
  </a>
  <span class="spacer"></span>
  <div class="links">
    <a href="/blog/">Latest</a>
    <a href="/blog/topics/">Topics</a>
    <a href="/blog/about/">About</a>
    <a href="https://munderdiffl.in">↗ Product</a>
  </div>
  <div class="nav-actions">
    <a class="btn sm" href="https://munderdiffl.in/blog/feed.xml" aria-label="RSS feed">⤵ RSS</a>
    <a class="btn sm primary" href="https://munderdiffl.in/#install">⤓ Download</a>
  </div>
</nav>

<main id="main">

<article class="t-guides">
  <header class="post-head wrap">
    <nav class="crumbs" aria-label="Breadcrumb">
      <a href="/blog/">Blog</a>
      <span>/</span>
      <a href="/blog/topics/guides/">Guides</a>
    </nav>
    <h1>Context Engineering for AI Agents: Budgeting the Window</h1>
    <p class="dek">Why context engineering beats prompt wording for agents — and the tactics (isolation, retrieval, externalized state, compaction) that keep the window lean.</p>
    <div class="byline">
      <span class="av" aria-hidden="true">CG</span>
      <strong>Chaitanya Giri</strong>
      <span class="dot" aria-hidden="true">·</span>
      <time datetime="2026-06-04">Jun 4, 2026</time>
      <span class="dot" aria-hidden="true">·</span>
      <span>7 min read</span>
      <span class="tag kind">Guides</span>
    </div>
  </header>

  <div class="hero-wrap wrap"><figure class="fig fig-hero"><img src="/blog/assets/media/context-engineering-for-ai-agents/hero.png" alt="Illustration: Context Engineering for AI Agents: Budgeting the Window" loading="lazy" decoding="async" /></figure></div>

  <div class="post-wrap wrap">
    <div class="prose">
      <div class="callout tldr"><span class="ic">TL;DR</span><p><strong>Context engineering</strong> is
deciding what occupies a model's finite window at each step — and what stays out. It's the system-level
successor to prompt engineering. The window is a <strong>budget</strong>, not a bucket: past a point,
more tokens make answers <em>worse</em>. The four moves that matter for a hive of agents are
<strong>isolation</strong> (a small window per agent), <strong>retrieval</strong> (pull the few
relevant notes, not all of them), <strong>externalized state</strong> (keep mailboxes, logs, and the
plan in files, not in context), and <strong>compaction</strong> (summarize the long tail). Get those
right and a fleet of agents stays sharp for hours.</p></div>
<p>For a long time the craft of working with language models was <em>prompt engineering</em> — finding the
wording that coaxed the best answer out of a single request. That framing made sense when the unit of
work was one prompt and one completion. It stops making sense the moment you have an agent that runs for
an hour, calls thirty tools, reads files, and coordinates with five other agents. At that scale the
exact phrasing of any one message is a rounding error. What decides whether the agent succeeds is the
<strong>entire set of tokens in its context window at the moment it has to act</strong> — and that set is something
you design. That design discipline is <strong>context engineering</strong>.</p>
<h2 id="the-window-is-a-budget-not-a-bucket" tabindex="-1">The window is a budget, not a bucket <a class="anchor" href="#the-window-is-a-budget-not-a-bucket" aria-hidden="true">#</a></h2>
<p>Every model has a finite context window, and that window is the agent’s entire working memory. Newer
models advertise enormous windows — hundreds of thousands of tokens, sometimes a million — and the
intuitive conclusion is “great, I’ll just put everything in.” That intuition is wrong, and it’s the
single most expensive mistake in agent design.</p>
<p>Quality does not increase monotonically with tokens. Past a certain fill level, answers get <em>worse</em> —
a phenomenon now widely called <strong>context rot</strong>. Irrelevant history distracts the model; stale tool
output gets mistaken for current state; two notes that contradict each other pull the model in two
directions. The window is a budget you spend, and every token you add has a cost in attention even
when it’s free in dollars. The job of context engineering is to spend that budget on the
<strong>smallest set of high-signal tokens</strong> that lets the model take the right next step.</p>
<p>It helps to name the ways a window goes bad. The 2025 taxonomy that stuck:</p>
<ul>
<li><strong>Poisoning</strong> — a wrong fact (often a hallucination) lands in context and gets referenced as truth on
every subsequent turn.</li>
<li><strong>Distraction</strong> — so much history accumulates that the model leans on it instead of reasoning about
the task in front of it.</li>
<li><strong>Confusion</strong> — irrelevant-but-present content (a tool the agent will never call, a doc for another
feature) nudges the model off course.</li>
<li><strong>Clash</strong> — two parts of the window disagree, and the model has no way to know which to trust.</li>
</ul>
<p>Notice that three of the four get <em>worse</em> as you add tokens. That’s the whole argument for treating
context as something to curate, not accumulate.</p>
<figure class="fig fig-inline"><img src="/blog/assets/media/context-engineering-for-ai-agents/note-1.png" alt="Hand-drawn sketch from “Context Engineering for AI Agents: Budgeting the Window”" loading="lazy" decoding="async" /></figure>
<h2 id="multi-agent-makes-it-sharper" tabindex="-1">Multi-agent makes it sharper <a class="anchor" href="#multi-agent-makes-it-sharper" aria-hidden="true">#</a></h2>
<p>In a single-agent setup you can sometimes get away with a messy window. In a hive you can’t. Every
agent has its own window, and they have to coordinate. The naive design — one giant shared context that
every agent sees — fails on all three axes at once: it’s expensive (you pay for the shared blob in every
agent’s token bill), it’s slow (bigger windows cost latency), and it’s confusing (each agent wades
through five other agents’ chatter to find its own task).</p>
<p>The whole architecture of <a href="https://munderdiffl.in/#what">a multi-agent harness</a> is, viewed one way, an
answer to a context-engineering question: <em>how do we let many agents share state without sharing
windows?</em> The answer is to keep the shared state <strong>out</strong> of the window and let each agent pull in only
the slice it needs.</p>
<h2 id="four-moves-that-keep-the-window-lean" tabindex="-1">Four moves that keep the window lean <a class="anchor" href="#four-moves-that-keep-the-window-lean" aria-hidden="true">#</a></h2>
<h3 id="1-isolation-one-small-window-per-agent" tabindex="-1">1. Isolation — one small window per agent <a class="anchor" href="#1-isolation-one-small-window-per-agent" aria-hidden="true">#</a></h3>
<p>The first and biggest lever is to give each agent its <strong>own</strong> context, scoped to its <strong>own</strong> role.
A planner doesn’t need the executor’s tool output; a reviewer doesn’t need the full chat history of the
agent it’s reviewing. When an orchestrator hands a sub-task to a specialist, it hands over a <em>brief</em> —
the goal and the few facts that matter — not a transcript.</p>
<p>This is why “spin up a sub-agent for the search” so often beats “stuff the search results into my own
context.” The sub-agent burns its window on the messy intermediate work and returns a clean conclusion;
the parent’s window stays small and high-signal. Isolation is context engineering’s version of
encapsulation. (For where this helps and where it doesn’t, see
<a href="/blog/claude-code-subagents-vs-multi-agent-harness/">Claude Code subagents vs a multi-agent harness</a>.)</p>
<h3 id="2-retrieval-recall-the-relevant-few-not-the-relevant-many" tabindex="-1">2. Retrieval — recall the relevant few, not the relevant many <a class="anchor" href="#2-retrieval-recall-the-relevant-few-not-the-relevant-many" aria-hidden="true">#</a></h3>
<p>An agent that has worked for weeks has learned far more than fits in a window. The wrong fix is to load
its whole history “just in case.” The right fix is <strong>retrieval</strong>: store everything durably outside the
window, and at each step pull in only the handful of notes that bear on the task.</p>
<p>This is exactly what <a href="/blog/semantic-memory-for-ai-agents/">semantic memory for AI agents</a> provides. The
agent’s knowledge lives in <a href="/blog/markdown-first-agent-memory/">markdown files</a>; a semantic index sits on
top; a query like <em>“how do we build the web bundle?”</em> returns the two or three relevant notes by
meaning, even if they never used those words. The window cost stays flat as the knowledge base grows —
that’s the property you’re buying. Recall-everything scales the window with history; retrieval keeps it
constant.</p>
<h3 id="3-externalized-state-files-the-agent-reads-not-tokens-it-carries" tabindex="-1">3. Externalized state — files the agent reads, not tokens it carries <a class="anchor" href="#3-externalized-state-files-the-agent-reads-not-tokens-it-carries" aria-hidden="true">#</a></h3>
<p>The most underrated context-engineering move is to keep coordination state in <strong>files</strong> rather than in
the window. A hive’s shared plan, its message mailboxes, and its event log are all state the agents need
<em>access</em> to — but not state they need <em>resident</em> in context at all times.</p>
<p>In a <a href="/blog/atomic-file-mailboxes-for-agents/">file-based coordination model</a>, an agent reads its inbox
at the start of a task, acts, and writes a reply to its outbox — the messages live on disk, not in a
shared prompt. The <a href="/blog/append-only-event-log-agents/">append-only event log</a> records every action for
debugging and replay, but no agent carries the log in its window; it’s read on demand when something
needs investigating. The shared board is the plan of record, edited by one writer, read by anyone who
needs it. The window holds a <em>pointer</em> to the state, and the agent dereferences it just-in-time. This is
the agent equivalent of not loading the entire database into RAM.</p>
<p>The same principle applies to tools and docs: load a tool’s full schema or a library’s documentation
<strong>when the agent decides to use it</strong>, not preemptively for every tool it might theoretically call.
Just-in-time beats just-in-case.</p>
<h3 id="4-compaction-summarize-the-long-tail" tabindex="-1">4. Compaction — summarize the long tail <a class="anchor" href="#4-compaction-summarize-the-long-tail" aria-hidden="true">#</a></h3>
<p>Even a well-scoped agent eventually fills its window if it runs long enough. The escape valve is
<strong>compaction</strong>: when the history grows past a threshold, summarize the old turns into a compact recap and
continue from there. Done well, the agent keeps its conclusions, its open decisions, and its current
state, while shedding the verbose middle — the tool calls and dead ends that got it there.</p>
<p>Our hive’s hourly <strong>standup</strong> is compaction as a team ritual: each agent reports what it did and what’s
next in a few lines, and that digest — not the raw hour of work — is what propagates. The transcript
stays on disk for anyone who needs to replay it; the <em>summary</em> is what travels. Whether it’s a single
agent compacting its own session or a team compacting into a standup, the move is the same: trade fidelity
you no longer need for window you do.</p>
<figure class="fig fig-inline"><img src="/blog/assets/media/context-engineering-for-ai-agents/note-2.png" alt="Hand-drawn sketch from “Context Engineering for AI Agents: Budgeting the Window”" loading="lazy" decoding="async" /></figure>
<h2 id="a-practical-checklist" tabindex="-1">A practical checklist <a class="anchor" href="#a-practical-checklist" aria-hidden="true">#</a></h2>
<p>When an agent underperforms, resist the urge to rewrite its prompt first. Audit its window instead:</p>
<ol>
<li><strong>What’s resident that shouldn’t be?</strong> Stale tool output, another agent’s chatter, a doc for a
feature this agent doesn’t touch — evict it.</li>
<li><strong>What’s missing that should be there?</strong> The one decision from three hours ago that changes the
answer. Make sure retrieval surfaces it.</li>
<li><strong>Is anything contradicting anything?</strong> Two notes that clash; an outdated fact next to its correction.
Resolve or remove.</li>
<li><strong>Could this be a sub-agent?</strong> If the task involves messy intermediate work, isolate it and return a
clean result.</li>
<li><strong>Is shared state in the window or in files?</strong> Move coordination state to disk and read it on demand.</li>
<li><strong>Is it time to compact?</strong> If history dominates the window, summarize and continue.</li>
</ol>
<p>Most “the model got dumber” complaints are context problems wearing a model costume. The fix is rarely a
better sentence; it’s a cleaner window.</p>
<p>The same care applies to what comes back out: <a href="/blog/asd-ste100-prompting/">ASD-STE100 prompting</a> is a one line way to make an agent’s answers shorter and easier to check.</p>
<h2 id="why-this-is-the-durable-skill" tabindex="-1">Why this is the durable skill <a class="anchor" href="#why-this-is-the-durable-skill" aria-hidden="true">#</a></h2>
<p>Models will keep getting bigger windows, and every jump tempts people to stop curating. Don’t. A bigger
window raises the ceiling on what you <em>can</em> hold; it does nothing to change the fact that signal beats
volume. The teams that get the most out of agents — single or swarm — are the ones who treat the context
window as the scarce, high-value resource it is and engineer what goes into it deliberately.</p>
<p>Munder Difflin is built around exactly these moves: <a href="https://munderdiffl.in/#what">isolated per-agent windows</a>,
<a href="https://munderdiffl.in/#how">semantic memory</a> that recalls the relevant few, file-based mailboxes and an
event log that keep shared state out of context, and standups that compact the long tail.
<a href="https://munderdiffl.in/#install">Download Munder Difflin</a> to run a hive that stays sharp for hours —
it’s free and open source.</p>


      
      <section class="faq" aria-label="Frequently asked questions">
        <h2 id="faq">FAQ</h2>
        
        <div class="qa">
          <p class="q">What is context engineering?</p>
          <p class="a">Context engineering is the discipline of deciding exactly what tokens occupy a model&#39;s finite context window at each step — instructions, tools, retrieved memory, and history — so the model has what it needs and nothing it doesn&#39;t. It&#39;s the system-level successor to prompt engineering: prompt engineering tunes the wording of one message, context engineering curates the whole window over a long-running, multi-turn agent.</p>
        </div>
        
        <div class="qa">
          <p class="q">How is context engineering different from prompt engineering?</p>
          <p class="a">Prompt engineering optimizes a single static prompt. Context engineering manages a dynamic, growing window across many turns and often many agents — what to load, what to retrieve just-in-time, what to summarize, and what to keep out of the window entirely. For agents that run for hours and call dozens of tools, the prompt is the smallest part of the problem.</p>
        </div>
        
        <div class="qa">
          <p class="q">What is context rot?</p>
          <p class="a">Context rot is the degradation in answer quality as a context window fills up. More tokens isn&#39;t strictly better: irrelevant history, stale tool output, and contradictory notes distract the model and crowd out what matters. A large window is a budget to spend carefully, not a bucket to fill.</p>
        </div>
        
        <div class="qa">
          <p class="q">Why does context engineering matter more in a multi-agent system?</p>
          <p class="a">Because every agent has its own finite window and they must coordinate. If you pour all shared state into every agent&#39;s context, you pay for it in tokens, latency, and confusion. The fix is to give each agent a small, role-scoped window and keep the shared state — mailboxes, logs, the plan — in files the agent reads on demand.</p>
        </div>
        
      </section>
      

      
      <hr />
      <div class="share">
        <span class="lbl">Tags</span>
        <a class="tag" href="/blog/tags/context-engineering/">Context Engineering</a><a class="tag" href="/blog/tags/memory/">Memory</a><a class="tag" href="/blog/tags/multi-agent/">Multi-Agent</a><a class="tag" href="/blog/tags/claude-code/">Claude Code</a>
      </div>
      
    </div>

    
    <aside class="toc-rail" aria-label="Table of contents">
      <div class="tt">On this page</div>
      <ol>
        <li class="lvl-2"><a href="#the-window-is-a-budget-not-a-bucket">The window is a budget, not a bucket</a></li><li class="lvl-2"><a href="#multi-agent-makes-it-sharper">Multi-agent makes it sharper</a></li><li class="lvl-2"><a href="#four-moves-that-keep-the-window-lean">Four moves that keep the window lean</a></li><li class="lvl-3"><a href="#1-isolation-one-small-window-per-agent">1. Isolation — one small window per agent</a></li><li class="lvl-3"><a href="#2-retrieval-recall-the-relevant-few-not-the-relevant-many">2. Retrieval — recall the relevant few, not the relevant many</a></li><li class="lvl-3"><a href="#3-externalized-state-files-the-agent-reads-not-tokens-it-carries">3. Externalized state — files the agent reads, not tokens it carries</a></li><li class="lvl-3"><a href="#4-compaction-summarize-the-long-tail">4. Compaction — summarize the long tail</a></li><li class="lvl-2"><a href="#a-practical-checklist">A practical checklist</a></li><li class="lvl-2"><a href="#why-this-is-the-durable-skill">Why this is the durable skill</a></li>
      </ol>
    </aside>
    
  </div>

  <footer class="post-foot wrap"><div class="prevnext">
      <a class="pn prev" href="/blog/debugging-multi-agent-systems/"><span class="k">← Older</span><span class="t">How to Debug a Multi-Agent System</span></a>
      <a class="pn next" href="/blog/claude-squad-vs-munder-difflin/"><span class="k">Newer →</span><span class="t">Claude Squad vs Munder Difflin: Which Should You Use?</span></a>
    </div>
    <section class="related">
      <h2>Related in Guides</h2>
      <div class="post-grid">
        <article class="card t-guides">
  <a class="thumb t-guides" href="/blog/claude-md-best-practices/" aria-hidden="true" tabindex="-1"><img src="/blog/assets/media/claude-md-best-practices/hero.png" alt="" loading="lazy" decoding="async" /></a>
  <div class="body">
    <div class="meta">
      <time datetime="2026-10-26">Oct 26, 2026</time>
      <span class="dot" aria-hidden="true"></span>
      <span>4 min</span>
    </div>
    <h3><a href="/blog/claude-md-best-practices/">CLAUDE.md Best Practices: What to Write, Where It Lives, How Long (Oct 2026)</a></h3>
    <p class="dek">CLAUDE.md best practices from Anthropic&#39;s docs, checked 26 Oct 2026: what to put in it, where each file lives, the 200 line target and how AGENTS.md loads.</p>
    <div class="foot">
      <span class="tag kind">Guides</span>
      <a class="more" href="/blog/claude-md-best-practices/" aria-label="Read: CLAUDE.md Best Practices: What to Write, Where It Lives, How Long (Oct 2026)">Read →</a>
    </div>
  </div>
</article>
<article class="card t-guides">
  <a class="thumb t-guides" href="/blog/codex-computer-use/" aria-hidden="true" tabindex="-1"><img src="/blog/assets/media/codex-computer-use/hero.png" alt="" loading="lazy" decoding="async" /></a>
  <div class="body">
    <div class="meta">
      <time datetime="2026-10-26">Oct 26, 2026</time>
      <span class="dot" aria-hidden="true"></span>
      <span>4 min</span>
    </div>
    <h3><a href="/blog/codex-computer-use/">Codex Computer Use: How to Turn It On, Plans and Limits (Oct 2026)</a></h3>
    <p class="dek">Codex computer use lets the ChatGPT desktop app see and click apps on macOS and Windows. Setup, which plans get it, locked use and how Claude&#39;s compares.</p>
    <div class="foot">
      <span class="tag kind">Guides</span>
      <a class="more" href="/blog/codex-computer-use/" aria-label="Read: Codex Computer Use: How to Turn It On, Plans and Limits (Oct 2026)">Read →</a>
    </div>
  </div>
</article>
<article class="card t-guides">
  <a class="thumb t-guides" href="/blog/cursor-pricing/" aria-hidden="true" tabindex="-1"><img src="/blog/assets/media/cursor-pricing/hero.png" alt="" loading="lazy" decoding="async" /></a>
  <div class="body">
    <div class="meta">
      <time datetime="2026-10-26">Oct 26, 2026</time>
      <span class="dot" aria-hidden="true"></span>
      <span>5 min</span>
    </div>
    <h3><a href="/blog/cursor-pricing/">Cursor Pricing (Oct 2026): Every Plan After the 12 Oct Change</a></h3>
    <p class="dek">Cursor pricing as of 26 Oct 2026: Free, Plus $8, Premium $30, Super $100, Ultra $200, Teams seats, what changed on 12 Oct and what old plans keep.</p>
    <div class="foot">
      <span class="tag kind">Guides</span>
      <a class="more" href="/blog/cursor-pricing/" aria-label="Read: Cursor Pricing (Oct 2026): Every Plan After the 12 Oct Change">Read →</a>
    </div>
  </div>
</article>

      </div>
    </section>
    
  </footer>
</article>

</main>
<footer class="site-foot">
  <div class="wrap">
    <div class="foot-top">
      <a class="foot-brand brand" href="https://munderdiffl.in" aria-label="Munder Difflin home">
        <img class="brand-logo" src="https://munderdiffl.in/logo.png" width="1318" height="840" alt="Munder Difflin, Inc. — Multi-Agent Harness" />
      </a>
      <div class="foot-links">
        <a href="/blog/">Latest</a>
        <a href="/blog/topics/">Topics</a>
        <a href="/blog/about/">About</a>
        <a href="https://munderdiffl.in/blog/feed.xml">RSS</a>
        <a href="https://github.com/chaitanyagiri/munder-difflin">GitHub</a>
        <a href="https://munderdiffl.in">munderdiffl.in ↗</a>
      </div>
    </div>
    <div class="foot-meta">
      © 2026 Munder Difflin · Local-first · MIT-licensed ·
      Built on <a href="https://munderdiffl.in">Electron, React, Pixi.js, xterm.js &amp; node-pty</a>.
      Run an office of agents while you sleep.
    </div>
  </div>
</footer>

<script>
  // sticky-nav border on scroll
  (function () {
    var nav = document.getElementById('nav');
    if (!nav) return;
    var onScroll = function () { nav.classList.toggle('scrolled', window.scrollY > 8); };
    onScroll(); window.addEventListener('scroll', onScroll, { passive: true });
  })();
  // YouTube click-to-load: thumbnail button → youtube-nocookie iframe
  (function () {
    document.addEventListener('click', function (e) {
      var btn = e.target.closest('.yt-load');
      if (!btn) return;
      var box = btn.closest('.yt');
      var id = box && box.getAttribute('data-yt');
      // Only an 11 character YouTube video id may reach the iframe URL.
      if (!id || !/^[A-Za-z0-9_-]{11}$/.test(id)) return;
      var frame = document.createElement('iframe');
      frame.src = 'https://www.youtube-nocookie.com/embed/' + encodeURIComponent(id) + '?autoplay=1';
      frame.title = box.getAttribute('data-title') || 'YouTube video';
      frame.allow = 'accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share';
      frame.setAttribute('allowfullscreen', '');
      box.replaceChildren(frame);
    });
  })();
</script>

</body>
</html>