本文由 AI 分析生成
建立時間： 2026-06-25 來源： https://x.com/0xwhrrari/status/2065539680146182270
Summary
Frames “loop engineering” as the successor skill to prompt engineering: instead of one prompt → one human-reviewed answer, you design a closed feedback system (discover → plan → execute → verify → iterate) that runs until a defined goal passes. Lays out six building blocks (automations, worktrees, skills, connectors, subagents, memory), the open-loop vs closed-loop distinction, and the token-cost realities that make most loops impractical without cheap long-context models.
將「循環工程（loop engineering）」定位為提示工程的下一階段技能：不再是單次提示換取人工審核的單一回答，而是設計一個封閉式回饋系統（發現→規劃→執行→驗證→迭代），持續運行直到達成設定目標。文章列出六個構成要素（自動化、worktree、技能、連接器、子代理、記憶），區分開放式與封閉式循環，並指出 token 成本是多數循環在缺乏便宜長上下文模型時難以實行的現實限制。
Key Points
- Loop engineering shifts the unit of work from “an instruction” to “a job”: goal, evaluator, stop condition, and persistence across runs.
- Closed loops (bounded, human-designed path, clear stop condition) are the practical starting point; open loops (exploratory, agent discovers the path) are powerful but expensive and prone to drift.
- The 6 building blocks: automations (trigger), worktrees (parallel isolation), skills (reusable context), connectors (real-tool access via MCP-like integrations), subagents (separate maker from checker), memory (state across runs).
- Single-agent loops suit focused/bounded tasks; fleet loops (orchestrator + specialists + subagents) suit larger multi-step goals like “build a productivity app.”
- Token economics are the real constraint: single tasks run 50K-200K tokens, fleets 500K-2M, scheduled daily loops reach millions/week — cheap, large-context, tool-calling models are a prerequisite, not a nice-to-have.
Insights
This piece and “Loops explained: Claude, GPT, Mira” (same batch) describe an almost identical five-block mental model from different angles, both citing the same Boris Cherny quote — suggesting this framing has become a a circulating meme/talking point in the Claude Code ecosystem rather than independently-derived advice. The maker/checker separation (subagents) recurs across nearly every source on loops as the single highest-leverage structural decision, more so than any tooling choice.
Connections
- Loops explained Claude, GPT, Mira and what actually works
- Loop engineering the 14-step roadmap from prompter to loop designer.
- The Three-Tier Agent Stack Boris Cherny Actually Runs
- loop-engineering
- ai-agent
Raw Excerpt
A prompt engineer says: “Write me a function.” A loop engineer says: “Write it, test it, fix it until it passes, then summarize the change.” Same tools. Different mindset.