The Three-Tier Agent Stack Boris Cherny Actually Runs

本文由 AI 分析生成

建立時間： 2026-06-19 來源： https://www.reddit.com/r/AIAgentsInAction/comments/1u2a583/the_threetier_agent_stack_boris_cherny_actually/

Summary

Describes the three-tier Claude Code agent stack attributed to Boris Cherny (head of Claude Code): Tier 1 /loop for session-scoped daytime cadences via slash commands, Tier 2 cloud-hosted Routines for overnight cron-style runs against fresh clones, and Tier 3 /batch/dynamic JavaScript workflows that fan a task out to hundreds of worktree-isolated subagents with a maker/adversarial-reviewer split. Includes a concrete 800-agent monorepo rename example and how the tiers feed each other via shared inbox/audit files.

描述歸於 Boris Cherny（Claude Code 負責人）的三層代理堆疊：第一層 /loop 為工作階段內、日間用的固定排程指令；第二層 Routines 為雲端託管、針對全新 clone 執行的夜間 cron 式排程；第三層 /batch 與動態 JavaScript 工作流，將任務拆分給數百個以 worktree 隔離的子代理，並配置製作者／對抗式審查者分工。文中以一個 800 代理規模的 monorepo 改名範例說明，以及三層如何透過共享的 inbox/audit 檔案互相串接。

Key Points

• Tier 1 /loop: session-scoped, min 1-minute interval, up to 50 active tasks, runs slash commands from .claude/commands/, can spawn a restricted-toolset subagent via --agent=.
• Tier 2 Routines: cloud-hosted against a fresh clone, no open session needed, min 1-hour interval, Anthropic adds up to 30 min of scheduling jitter.
• Tier 3 /batch and dynamic workflows: JavaScript harnesses using agent(), parallel(), pipeline() that Claude writes on the fly; each spawned agent gets its own worktree and is adversarially reviewed by a second agent instance before results synthesize.
• The tiers connect through shared state: Routines write findings to .claude/audit/ or .claude/inbox/; daytime loops read and act on that; oversized jobs get escalated to /batch or a dynamic workflow.
• A real cited example: an 800-agent swarm migrating every user.email callsite to user.primaryEmail across a 4,000-file monorepo, each agent in its own worktree, running tests and a second-agent adversarial review per file.
• A top comment raises a real concern (verified against the source thread, not vault-attributed fact): periodic CLAUDE.md “distiller” loops that only ever append rules compound a per-token cost across every overnight spawn, since the file is reread on every fresh clone.

Insights

The Reddit comments are part of the source material and contain unverified/skeptical claims (one accuses Cherny of inflating token spend for IPO optics) — flagged here as the kind of content that should be weighted as discussion, not fact. The CLAUDE.md-distiller cost-compounding comment is the most concretely actionable critique in the whole batch of loop-engineering clippings: a rule added once gets paid for on every single spawn forever, which argues for periodically auditing distilled rules for actual hit-rate, not just appending.

Connections

• Loops explained Claude, GPT, Mira and what actually works
• Loop Engineering The AI skill every builder needs in 2026
• Loop engineering the 14-step roadmap from prompter to loop designer.
• loop-engineering
• claude-code

Raw Excerpt

Routines write structured output to .claude/audit/ or .claude/inbox/. Loops in your morning session read from there and act on it. … The compound effect is in the system, not the model.