Multi-agent research pipelines for Claude Code — stop leaving your session to go research elsewhere.
You're mid-session, you need real research, and you don't want to break flow to Perplexity or ChatGPT and paste results back. Or you want to deeply understand an open-source repo before building on it. Or you need structured data on 20 competitors in a consistent schema.
These pipelines delegate research to agent teams so your top-level Claude stays free. Results come back as committed markdown in docs/research/ — artifacts, not chat messages.
This is a Claude Code plugin. The first-class install path is to hand the job to your agent — paste the prompt below into a Claude Code session and let it follow the playbook in docs/install.md.
Please install the deep-research-claude plugin from https://github.com/dbc-oduffy/deep-research-claude. The repo's
docs/install.mdis the install playbook — read it and follow the steps. Verify by running/deep-research setupat the end and report the result.
After the agent finishes, restart Claude Code (so the Agent Teams env var takes effect) and try:
/deep-research:research --mode=web "agent orchestration patterns in LLM frameworks"
Haiku scout builds a source corpus via web search. 3-5 Sonnet specialists deep-read sources, verify claims, and challenge each other adversarially. Opus sweep agent checks coverage, fills gaps, and writes the final document. An optional iterative deepening pass targets high-severity gaps identified in the first sweep.
/deep-research:research --mode=web "topic"
2 Haiku scouts inventory every file in assigned chunks. 4 Sonnet specialists deep-read, analyze architecture, and optionally compare against a second project. Opus synthesizer writes the final assessment.
/deep-research:research --mode=repo /path/to/repo [--compare /path/to/mine] [--survey] [--deeper] [--deepest]
--compare— gap-analysis artifact comparing target repo to your project--deeper— dependency-weighted repomap during scoping; specialists prioritize structurally central files--deepest— adds a Sonnet atlas agent producing architecture artifacts (file index, system map, connectivity matrix)
Schema-conforming batch research across N entities. Haiku scout maps findings to schema fields. 1-5 Sonnet verifiers challenge each other's values (CONFIRMED / UPDATED / REFUTED / CONTESTED). Opus synthesizer resolves contested fields and outputs validated YAML/JSON.
/deep-research:research --mode=structured tasks/research/spec.yaml subject-key
Research YouTube videos, podcasts, and media Claude can't access directly, via NotebookLM. Haiku scout ingests sources. Sonnet workers query on focused sub-questions. Opus sweep writes the synthesis.
Requires the notebooklm-mcp-cli MCP server and a Google account with NotebookLM access.
/notebooklm-research "topic"
| Command | Pipeline |
|---|---|
/deep-research:research --mode=web <topic> |
Internet research with iterative deepening |
/deep-research:research --mode=repo <path> |
Repository analysis (with optional --compare, --survey, --deeper, --deepest) |
/deep-research:research --mode=structured <spec> <key> |
Schema-conforming batch research |
/notebooklm-research <topic> |
NotebookLM media research |
All pipelines are fire-and-forget — the EM spawns the team and is freed. Results are committed to docs/research/ automatically.
| Agent | Model | Role |
|---|---|---|
| research-scout | Haiku | Web search, source vetting, shared corpus |
| research-specialist | Sonnet | Source verification, adversarial peer challenges, structured claims |
| research-synthesizer | Opus | Coverage check, gap-filling, final document |
| repo-scout | Haiku | File inventory with signatures, constants, data flow |
| repo-specialist | Sonnet | Architecture analysis, optional project comparison |
| structured-synthesizer | Opus | Schema validation, contested field resolution, final YAML/JSON |
This plugin works standalone. When used alongside the coordinator plugin, a PreToolUse hook automatically suggests research pipelines when Claude reaches for ad-hoc web search — nudging toward these structured pipelines instead of one-off WebFetch calls that consume the coordinator's context window.
Pipeline commands (/web, /repo, /structured) invoke ~/.claude/plugins/coordinator/bin/coordinator-safe-commit for phase-end commits. The helper provides scoped staging (per-session audit-trail integrity) for users running multiple concurrent agent sessions on the same branch.
If you have the coordinator plugin installed, no action needed — the helper is on the expected path.
If running deep-research standalone, substitute either form when you encounter the command:
git add <explicit-paths> && git commit -m "<subject>"— manual scoped staginggit add -A && git commit -m "<subject>"— blanket staging (acceptable for solo single-session use)
Pipeline design derives from published guidance (OpenAI, Perplexity, Google, Anthropic, Stanford STORM) and is validated through controlled experiments. Anthropic independently built a production multi-agent research system using the same core pattern — their eval showed 90.2% improvement over single-agent. We converged on the same architecture independently; this system extends it with Haiku scouts for cost efficiency, adversarial peer dynamics between specialists, and asynchronous orchestrator dispatch.
This is the canonical home of the deep-research plugin. Originally developed as part of coordinator-claude and extracted for independent distribution.
Pipeline D is built on notebooklm-mcp-cli by jacob-bd — an MCP server that provides programmatic access to Google NotebookLM.
the PM O'Duffy & Claude