Mapping a legacy codebase with 16 parallel reverse-engineering agents
A Claude Code swarm trick: Opus splits 578 files, 16 agents reverse-engineer one service each and return only summaries, while Jev routes micro-decisions in milliseconds.
THIS IS ACTUAL F*CKING GOLD Anthropic just shared one of the best tricks for Claude Code. The idea: before you rewrite a legacy codebase, map it with 16 reverse-engineers at once and never let the big model touch the old code. 578 files, 16 services, mapped in 4 minutes. Opus alone was 6% through and already sitting on 34.2k of context. You just launch Claude Code like this: claude --model opus --rea-swarm 16 legacy/ here's the split: → Opus 5.5 splits 578 files into 16 services, one per agent → 16 REA agents reverse-engineer a service each, in their own context → every agent returns a summary, never the code it read → Opus 5.5 merges 16 summaries into one system-map.md the part everyone gets wrong: they let the expensive model read the repo. alone, Opus would have to read 578 files, 412k of context, and it drowns long before it understands the thing. here it reads 16 summaries. 8.6k of context total, and it still ends up with the full map: services, rules, edge cases, conflicts. the agents did the reading, Opus did the understanding. the raw code never left the swarm. Opus splits. REA reads. Opus merges. JEV engineering is the same move one layer down: the forks that need no thinker (which file next, which tool, retry or stop) go to JEV in milliseconds, and the big model only sees the ones that actually split the full tree: > Opus 5.5 runs the main session and splits the repo into services > 16 REA agents reverse-engineer one service each, in parallel > every agent reads its own files and returns only a summary > raw code never enters the main context > Opus 5.5 merges 16 summaries into one system map with rules and conflicts > 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 and split any legacy repo into services before reading anything. 2. Spawn one REA agent per service, each with its own context, all in parallel. 3. Every agent returns a structured summary only. Raw file contents never enter the main session. 4. Merge the summaries into a single system map listing endpoints, rules, edge cases and conflicts between services. 5. Add one rule to CLAUDE.md: no rewrite starts until the map exists and every conflict is resolved. 6. Route fast deterministic forks (file order, 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, you'll use this on the first legacy repo you touch.
This post quotes another post: x.com/sprytixl/status/2107763347097534680