用 Opus 5.5 一条命令生成 16 个 REA 逆向工程 agent

Sprytix @Sprytixl原推发布于 2026.10.19

在 X 上打开原推

THIS IS A PURE F*CKING GOLDMINE Anthropic just shared one of the best tricks for Claude Code. The idea: you stop writing agent configs by hand. One prompt, and Opus 5.5 writes the entire swarm for you. By hand that is 16 prompts, 16 tool configs, 16 test harnesses and 2,312 lines you type yourself. Here you type one line. You just launch Claude Code like this: claude --model opus "build REA swarm ×16 for legacy/" here's the split: → Opus 5.5 reads 578 files and splits them into 16 services → it writes 16 specs in parallel, 1,037 lines, no templates → every spec becomes one REA agent, one reverse-engineer per service → each agent gets tested as it lands, 3/48 checks already green at 3 minutes the part everyone gets wrong: they hand-write every agent and then wonder why 16 of them drift apart. here one architect writes all 16 specs from the same read of the codebase, so every agent is wired the same way and nothing gets forgotten. REA is the piece doing the real work underneath: each agent reverse-engineers its own service, dry-runs on three files first, then writes its spec from what it actually found, not from what you guessed. you wrote 1 line. Opus wrote 1,037. Opus architects. REA reverse-engineers. The swarm builds itself. JEV engineering is the same move one layer down: the forks that need no thinker (which service, which tool, retry or stop) go to JEV in milliseconds, and the architect only sees the ones that actually split the full tree > Opus 5.5 runs as architect, reads the repo and splits it into services > it writes one spec per service, all 16 in parallel, from a single read > every spec spawns one REA agent that reverse-engineers that service > each agent dry-runs on a few files before writing anything > specs get tested as they land, the architect monitors all 16 in flight > Opus 5.5 deploys once every spec is ready JEV routes micro-forks in milliseconds paste the tree and this prompt into Claude Code ↓ "Rebuild my Claude Code setup around this tree: 1. Run the main session on Opus 5.5 as an architect that writes agent specs, never agent code by hand. 2. Point it at a repo, let it split the repo into services and write one spec per service, all in parallel from a single read. 3. Every spec spawns one REA agent that reverse-engineers its own service and dry-runs on a few files before writing anything. 4. Test each spec as it lands, never wait for all of them to finish first. 5. Add one rule to CLAUDE.md: no agent ships until its spec passes its own checks. 6. Route deterministic forks (service choice, tool choice, retries) through JEV. Show me every change as a diff first. No edits until I say go." ↳ http://code.claude.com/docs/en/model-config save it, it'll come in handy.

这条是引用着另一条写的：x.com/sprytixl/status/2107763347097534680