Back to blog
Playbook·7 min read

Research Workspace: Two Agents, One Source of Truth

Run a research loop with two agents and shared notes so sources live in one place. Do not let two models invent conflicting citations.

By the flAim team

A research workspace is two agents and one notes doc for sources. The notes are the source of truth. The models are staff. If the notes and the answers disagree, the notes win until you verify.

The failure mode is specific: two fluent models invent two citation lists. You merge them in a hurry. A claim that never had a source ships.

Setup

  1. Create a workspace from the Research template, or add two agents and a Notes surface.
  2. Paste the question, audience, and “must cite / must not guess” rules into notes first.
  3. Agent A: gather claims and candidate sources into the notes, not only into chat.
  4. Agent B: attack the notes. Flag unsourced claims. Do not add new facts without a source line.
  5. You (or a teammate) confirm or delete. Only then draft.

The citation rule

If you seeDoDo not
Two URLs that disagreeKeep both in notes with a one-line conflictPick the more confident-sounding model
A claim with no source lineDelete it from the notes or mark UNVERIFIEDLet it into the outline “for now”
A model offering a DOI-looking stringOpen it. If it 404s, it is not a sourceAssume academic tone equals a real paper
Chat disagreementTreat it as a signal. Score against the notesAverage the two answers into a third fiction

Chat vs Build for research

Use Chat when you want to steer: @ agent A to gather, @ agent B to attack the notes (it already saw A), then send to both. Use Build when the chain is gather → critique → outline and you do not want to run that playbook by hand. Inspect the plan. The orchestrator is reasonable routing, not a librarian.

Shared notes how-to: Documentation. Next playbook in this series: writing studio.

Start from the Research Workspace template

Two research agents plus shared Notes for sources. Free credits, no API key.

Related: Shared context · Comparison method · Docs

We're building this in the open and we read everything. If you try it and something's clumsy, tell us. We'd rather hear it from you than not hear it.