<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Minglong Pan — The Writings</title>
    <atom:link href="https://www.minglongpan.com/rss.xml" rel="self" type="application/rss+xml" />
    <link>https://www.minglongpan.com/#writing</link>
    <description>Field notes on production AI infrastructure, AI agents, deterministic LLM workflows, and human attention. A living document, maintained by the fleet.</description>
    <language>en-us</language>
    <lastBuildDate>Thu, 20 Aug 2026 00:00:00 GMT</lastBuildDate>
    <item>
      <title>A coherent virtual pet without an LLM</title>
      <link>https://www.minglongpan.com/writing/a-coherent-virtual-pet-without-an-llm</link>
      <guid isPermaLink="true">https://www.minglongpan.com/writing/a-coherent-virtual-pet-without-an-llm</guid>
      <pubDate>Thu, 20 Aug 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>Coherence in an ambient pet comes from explicit state, bounded choices, and reliable fallbacks — not from generating every line at runtime. The pattern: a two-axis mood lattice, an authored phrasebook with a strict fallback order, event-indexed selection, a conservative intent classifier, and a validator-gated LLM lane that can only ever audition.</description>
    </item>
    <item>
      <title>Why tmux is the wrong IPC layer for an agent fleet</title>
      <link>https://www.minglongpan.com/writing/why-tmux-is-the-wrong-ipc-layer-for-an-agent-fleet</link>
      <guid isPermaLink="true">https://www.minglongpan.com/writing/why-tmux-is-the-wrong-ipc-layer-for-an-agent-fleet</guid>
      <pubDate>Mon, 10 Aug 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>Ten times in eighteen days I pasted one AI agent's output into another agent's window. There is no wire between them. The wire is me.</description>
    </item>
    <item>
      <title>Why a self-improving agent's accept rate isn't a quality metric</title>
      <link>https://www.minglongpan.com/writing/accept-rate-is-not-a-quality-metric</link>
      <guid isPermaLink="true">https://www.minglongpan.com/writing/accept-rate-is-not-a-quality-metric</guid>
      <pubDate>Sat, 08 Aug 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>My eval gate scored a byte-identical file 0.167 better than itself and approved it. The number I trusted for months only recorded my own approvals.</description>
    </item>
    <item>
      <title>finance-os: a decision system, not a stock picker</title>
      <link>https://www.minglongpan.com/writing/finance-os-design-notes</link>
      <guid isPermaLink="true">https://www.minglongpan.com/writing/finance-os-design-notes</guid>
      <pubDate>Fri, 24 Jul 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>A personal investment platform that scores a ~70-name universe on 11 LLM-graded dimensions, blends probability-weighted scenarios, and treats the kill trigger as a first-class object — the design notes.</description>
    </item>
    <item>
      <title>How I run a personal agent fleet</title>
      <link>https://www.minglongpan.com/writing/how-i-run-a-personal-agent-fleet</link>
      <guid isPermaLink="true">https://www.minglongpan.com/writing/how-i-run-a-personal-agent-fleet</guid>
      <pubDate>Fri, 24 Jul 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>Runtime-agnostic agents on Telegram — Claude Code, Hermes, Codex, Kimi as interchangeable labor — with a shared knowledge memory, cron-driven loops, and a rule that everything the fleet ships gets verified before it's claimed.</description>
    </item>
    <item>
      <title>Persona-routed agent tooling: one stdio server, every team's capabilities</title>
      <link>https://www.minglongpan.com/writing/mcp-resource-layering-pattern</link>
      <guid isPermaLink="true">https://www.minglongpan.com/writing/mcp-resource-layering-pattern</guid>
      <pubDate>Fri, 24 Jul 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>A semantic router that reads the task or todo list, routes to persona → skills → tools, and auto-loads team resource packages — one stdio binary that works with any CLI or agentic tooling, with team behavior living in its own JSON config.</description>
    </item>
    <item>
      <title>Valhalla: an attention layer for a fleet of agents</title>
      <link>https://www.minglongpan.com/writing/valhalla-design-notes</link>
      <guid isPermaLink="true">https://www.minglongpan.com/writing/valhalla-design-notes</guid>
      <pubDate>Fri, 24 Jul 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>The operations layer for my agent fleet — agents post only what needs a decision to one attention feed, over a shared knowledge memory. Design notes on exception-driven surfaces and the four friction laws.</description>
    </item>
    <item>
      <title>Three agent frameworks converged on a control-plane protocol this month. I run my fleet on tmux and chose not to adopt it.</title>
      <link>https://www.minglongpan.com/writing/three-agent-frameworks-converged-on-a-control-plane-protocol</link>
      <guid isPermaLink="true">https://www.minglongpan.com/writing/three-agent-frameworks-converged-on-a-control-plane-protocol</guid>
      <pubDate>Mon, 20 Jul 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>Three frameworks shipped the same control-plane idea in a month — external process control, capability negotiation, env-var hygiene. What converged, why my tmux fleet already has the failure modes it solves priced in, and the two concrete triggers that would make me adopt it.</description>
    </item>
    <item>
      <title>Anti-Engagement AI: Software You Can Walk Away From</title>
      <link>https://minglong1.substack.com/p/anti-engagement-ai-software-you-can</link>
      <guid isPermaLink="true">https://minglong1.substack.com/p/anti-engagement-ai-software-you-can</guid>
      <pubDate>Mon, 01 Jun 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>As AI makes infinite engagement easier, the better product bet may be software that respects absence.</description>
    </item>
    <item>
      <title>ThreadLang: a deterministic DSL for LLM workflows</title>
      <link>https://www.minglongpan.com/writing/threadlang</link>
      <guid isPermaLink="true">https://www.minglongpan.com/writing/threadlang</guid>
      <pubDate>Mon, 01 Jun 2026 00:00:00 GMT</pubDate>
      <dc:creator>Minglong Pan</dc:creator>
      <description>Most multi-step LLM code is prompt strings glued together in a Python function. ThreadLang makes the workflow the program — a grammar, an AST, and a structured trace you can run without a network call.</description>
    </item>
  </channel>
</rss>
