Agent research lab

Research lab of agents

One agent researching a question returns whatever its first searches surfaced, citations included, some of them invented. A lab splits the roles that keep research honest: searchers cover the question in parallel, a verifier re-checks every source, and a synthesizer writes the answer with each claim attributed. The verifier is the point: an unchecked agent citation is a guess wearing a URL.

Team

• Lead (research-synthesis): sharpens the question and splits it.
• Searchers: fan out on sub-questions in parallel.
• Verifier: re-fetches each source and confirms the claim.
• Synthesizer (research-synthesis): writes the cited answer.

Shape: parallel fan-out, then a verification gate, then synthesis; see the deep-research harness.

Method

1. Sharpen and decompose the question. The lead turns "what about X" into a precise question, then splits it into non-overlapping sub-questions, one per searcher, so fan-out does not duplicate work.
2. Fan searchers out in parallel. Each returns sources.jsonl records: claim, url, publication date, exact quote, and the source's incentive. Primary sources (spec, paper, changelog) outrank commentary about them.
3. Assign one agent to the negation. Someone searches for disconfirming evidence explicitly, so the lab is not just a confirmation machine for the first hypothesis.
4. Verify every source independently. A separate agent re-fetches each url, confirms the quote exists and supports the claim, and marks it confirmed, refuted, or unsupported. Unsupported citations are cut, not softened.
5. Reconcile conflicts on the record. Where sources disagree, adjudicate by recency and method, or declare the point genuinely unsettled. A synthesis that hides its conflicts is an advertisement.
6. Synthesize in three layers. report.md gives the answer in two or three sentences, then the attributed evidence, then the edges: what is uncertain and what would settle it. Mark every inference as inference.

Run it

In Claude Code, spawn searchers as parallel subagents in one turn, each with a sub-question and web access, writing to a shared sources.jsonl; then run the verifier as its own subagent over that file before the synthesizer reads only verified rows. Port it to CrewAI as parallel search tasks feeding a synthesis task, to AutoGen as a GroupChat with a dedicated fact-checker, or to LangGraph as a fan-out then fan-in graph with a verification node between search and synthesis.

Signals it works

• Every load-bearing claim in the report cites a source the verifier confirmed.
• At least one searched-for counter-source appears, or its absence is stated.
• The top two or three sentences survive a skeptic re-checking their citations.

Boundaries

This organizes agent research; it does not make agents trustworthy on paywalled, very recent, or specialist sources they cannot actually fetch. Verification catches fake citations, not deep methodological flaws in a real paper, which needs human expertise. Treat the output as a sourced draft for a human to sign, not a final authority.