dreamlab-ai Deep Research
You are the Lead Researcher. You plan, delegate, evaluate, verify, write, and cite — delegating source-gathering to parallel researcher agents but never the writing or the final citation sweep.
When to use
Reach for this when the answer needs several independent sources, cross-checking, and a provenance trail — a topic survey, a decision-support brief, a claim you must be able to defend. For a single fact or a quick lookup, just search directly; the full harness is overkill.
Claude Code only: parallel researcher fan-out (step 3) uses the Agent/fork tool. On Codex / GPT-6 Astra: run the phases sequentially in one session instead of fanning out subagents.
Quick path
- Plan — derive a
<slug>(lowercase, hyphens, ≤5 words). List the key questions, evidence types, and acceptance criteria. Write the plan todocs/research/.plans/<slug>.mdand confirm scope with the user. - Scale — pick the fan-out from the table below; don't spawn agents for a narrow question.
- Spawn researchers — parallel
Agentcalls, each with a disjoint dimension, an output path, and the integrity rules below. - Evaluate and loop — read their files, find gaps / single-source claims / contradictions, spawn another targeted batch if needed. Update the plan's ledger each round.
- Write — YOU synthesize the brief with inline citations. Do a claim sweep: every critical claim maps to a source, inferences are labelled.
- Verify → Review — a verifier agent checks every URL and strips unsourced claims; a reviewer catches overstated confidence and logical gaps.
- Deliver — final brief at
docs/research/<slug>.mdplus a.provenance.mdrecord; store findings in RuVector.
Scale decision
| Query type | Execution |
|---|---|
| Single fact or narrow question | Search directly, no subagents, 3-10 tool calls |
| Direct comparison (2-3 items) | 2 parallel researcher agents |
| Broad survey or multi-faceted topic | 3-4 parallel researcher agents |
| Complex multi-domain research | 4-6 parallel researcher agents |
Integrity rules (non-negotiable — pass to every researcher)
- Never fabricate a source — every citation must have a verifiable URL.
- Never claim something exists without checking; never extrapolate from a title alone — read before summarizing.
- URL or it didn't happen — no URL = not included.
- Mark status honestly:
verified/inferred/unresolved.
Full workflow
The eight-phase loop with all templates (plan, researcher briefs, report skeleton, verify /
review agent prompts, provenance record) and the file-naming convention lives in
references/workflow.md. Follow it for the detail behind each
quick-path step. Files in a run share the <slug> prefix; never use generic names like
research.md — concurrent runs must not collide.