本文由 AI 分析生成
建立時間： 2026-07-30 來源： https://x.com/phosphenq/status/2068468824421364103
Summary
@phosphenq synthesizes Boris Cherny’s 259-PR/month autonomous coding workflow and the Claude Code ecosystem’s loop primitives: State, Automations, Worktrees, Skills, Connectors, and Sub-agents. The thread frames autonomous coding loops as infrastructure requiring deliberate design, using a $47K runaway loop as the cautionary tale and the maker/checker split as the core safety pattern.
@phosphenq 綜合了 Boris Cherny 每月 259 個 PR 的自主程式碼工作流程，以及 Claude Code 生態系的迴圈基礎元件：State、Automations、Worktrees、Skills、Connectors、Sub-agents。以 $47K 的失控迴圈作為反面教材，maker/checker 分離作為核心安全模式，將自主程式碼迴圈定位為需要刻意設計的基礎設施。
Key Points
- Boris Cherny: 259 PRs in December 2025, every one by Claude, editor never opened; $47K runaway loop documented as cautionary tale
- 6 infrastructure pieces: State (STATUS.md), Automations (/loop + /goal), Worktrees (per-agent isolation), Skills (codified intent per run), Connectors (MCP), Sub-agents (maker/checker split)
- /goal command: runs until a condition is true, graded by a separate smaller model; example: “all tests in test/auth pass and the lint step is clean”
- Estimated cost for 100-agent fleet: ~$1.3M/month (sponsored)
- Four loop deaths: runaway recursion, silent death (context full), random walk (no fixpoint), comprehension debt
- Core scoping principle: “Scope a loop by what it can destroy, not by what you want it to do”
Insights
The /goal command (condition-based termination graded by a separate model) is more robust than MAX_ITER alone because it aligns loop termination with actual success rather than with effort count. The maker/checker split combined with a different grading model prevents the agent from “teaching to the test”—finding shortcuts that satisfy the same model that wrote them. The $47K runaway example makes the isolation/brakes discussion concrete in a way that abstract risk warnings do not. The scoping principle (“what it can destroy”) is a better mental model than “what you want it to do” for determining loop blast radius.
Connections
- Loop Engineering A Technical Roadmap for an Autonomous Loop — Technical implementation complement with detailed step-by-step infrastructure
- How to Build a Claude Code Agent Team That Runs in Loops (Exact Setup Inside) — Simpler concrete 3-file implementation
- claude-code
- autonomous-agents
Raw Excerpt
“Last December, Boris Cherny shipped 259 pull requests in a single month. Every one written by Claude. He says he did not open an editor the whole time.”