AGENTS.md — .claude/
Claude Code internal subagents. These are NOT end-user agents — they're tool-using sub-processes the Claude Code harness spawns to parallelize work or protect the main context window.
Subagents available
| Name | Purpose | Tools |
|---|---|---|
w1-recon | Read files/docs verbatim, no decisions | Read, Grep, Glob, Bash |
w2-decide | Take W1 findings → produce structured plan + rubric | Read, Grep, Glob, Write |
w3-edit | Execute edits from W2 spec with exact anchors | Read, Edit, Write, Grep, Glob, Bash |
w4-verify | Run biome/tsc/vitest, score rubric, pass/fail gate | Read, Grep, Glob, Bash, Edit |
Explore | Fast codebase exploration | All except Agent/Edit/Write |
Plan | Implementation plan synthesis | All except Agent/Edit/Write |
general-purpose | Open-ended research + multi-step tasks | All |
See agents/ subdirectory for the full system prompts.
Why this folder, not agents/?
The sibling top-level agents/ folder is for substrate-deployed agents —
markdown definitions parsed into TypeDB, wired into the runtime, addressable
by channels (Telegram, Discord, web). These internal subagents live in
Claude Code only — they never reach production, never get a UID in the
substrate, never participate in pheromone routing.
Different audience. Different lifecycle. Same vocabulary (W1-W4 waves map 1:1 to the /do cycle).
How to invoke
From within Claude Code:
Agent(subagent_type="w2-decide", prompt="...", description="...")
Or via slash commands that invoke them transparently:
/do TODO-feature.md # spawns w1 → w2 → w3 → w4 in sequence
/do TODO-feature.md --auto # runs full cycle unattended
See also
- Root
AGENTS.md— substrate-deployed agents CLAUDE.md— harness compositioncommands/do.md— the W1-W4 cycle runner