deep-research

Deep Research

Create high-fidelity research reports with strict format control, evidence mapping, source governance, and multi-pass synthesis.

Architecture: Lead Agent + Subagents

Lead Agent (coordinator — minimizes raw search context)
  |
  P0: Environment + source policy setup
  |
  P1: Question and claim map (decision questions, evidence routes, stop rules)
  |
  Dispatch ──→ Subagent A ──→ writes task-a.md ──┐
           ──→ Subagent B ──→ writes task-b.md ──┤ (parallel)
           ──→ Subagent C ──→ writes task-c.md ──┘
  |                                               |
  |     research-notes/  <────────────────────────┘
  |
  P2: Dispatch + investigate → evidence packets
  P3: Citation registry + source governance
  P4: Evidence-mapped outline with counter-evidence and unknowns
  P5: Draft from evidence packets; reopen decisive originals
  P6: Counter-review (claims, confidence, alternatives)
  P7: Verify → asset check and catalog registration

Context discipline: Keep raw search-result noise in task workspaces. Pass evidence packets to the lead agent, including locators and short source excerpts. Notes are routing aids, not authorities: the lead agent must open the original source for every load-bearing claim, conflicting claim, and exact figure/date/quotation used in the report.

Run and asset contract — every invocation

Before external retrieval, read research-asset-contract.md. Create or resume a durable project study for single-route and multi-route research alike. Search its explicit prior-study catalog, open relevant earlier originals and record reuse/adapt/reject decisions. Draft the decision questions and provider × actual-mode lanes in study.json. Write one exact dispatch_context with the user-supplied seed URL, named entities, and verified codes or other identifiers needed for the query. Include that string in every lane prompt; research_assets.py start rejects an omission. Capture the user's named seed article or document as an original before external dispatch, and follow the provider handoff when a provider cannot read its URL. A model workspace may not see the user's link or local files just because the coordinator does. Then run provider_runs.py plan. Keep every user-requested mode in the plan; a route that adds no value, is unavailable or lacks paid authorization gets a reasoned deferred event. A report assembled from direct web search and internal subagents still needs a direct-source lane and the same archive.