Multi-agent research and investigation. Spawns a team of investigators to explore different facets of a topic in parallel. One is always Codex (different model family) for cross-validation.
Inspired by Anthropic's multi-agent research system.
/research Technical spike: Japanese character support impact assessment
/research size:5 depth:comprehensive Should we migrate from REST to GraphQL?
/research Do we need offline indicators on our Apple Watch app?| Arg | Example | Description |
|---|---|---|
size:N |
size:5 |
Team size (3-7) |
depth:X |
depth:comprehensive |
focused (1-2 rounds), standard (2-3), comprehensive (4+) |
| (rest) | Japanese character support |
Research topic (required) |
- Context — explores codebase (or skips for non-code topics), asks high-impact questions
- Scope — scores topic on 4 dimensions (4-12), sets team size and tool call budget
- Investigators — decomposes topic into facets, assigns domain-specific investigators (e.g., "Storage Engineer", "Filesystem Specialist", "Search Infrastructure Engineer")
- Team — spawns one agent per investigator, all working in parallel
- Research loop — investigate -> evaluate -> cross-pollinate -> evaluate -> deep-dive if needed
- Verification — Integrator validates all findings against cited sources
- Report — key findings with evidence tags, risk matrix, unknowns map, recommendations
- Orchestrator-worker pattern — lead agent coordinates independent investigators
- Codex scales with team — 1-3 Codex-powered investigators depending on team size, never more than half
- Facet-based decomposition — each investigator has clear boundaries (what's in/out of scope)
- Adaptive rounds — lead agent decides if more investigation is needed, not fixed round count
- Evidence tagging — findings tagged [grounded], [informed], or [speculative]
- Start wide, then narrow — investigators begin with broad searches, then focus
- Verification step — Integrator validates citations before final report
- Mandatory question enforcement — agent questions surfaced to user via AskUserQuestion
| /debate | /research | |
|---|---|---|
| Goal | Decide between options | Map unknowns, assess impact |
| Method | Hegelian dialectic (thesis/antithesis/synthesis) | Investigate/cross-pollinate/consolidate |
| Roles | Contrarian + Synthesizer | Scope Sentinel + Integrator |
| Output | Decision + consensus labels | Findings + risk matrix + unknowns map |
- Codex CLI installed
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1in Claude Code settings
MIT