---
name: multi-agent-ux-research
description: Plan and synthesize UX desk research using independent research tracks and a shared evidence ledger. Use for product research briefs with multiple questions or source types.
---

# Multi-agent UX research

Draft v0.85 by Mukund Sharma. The workflow and evidence contract are written; live calibration and repeatability evaluation remain in progress.

## Frame the brief

Identify the decision the research should support, the audience, the product stage, the timeframe and available evidence. If a missing detail changes the research direction, ask one focused question; otherwise record the assumption and continue. Distinguish desk research from interviews and usability testing.

## Split independent questions

Delegate bounded tracks when parallel work adds value: product context, comparable workflows, and public user signals. Give each track a research question, source scope and output contract. If agents are unavailable, run those tracks sequentially. Avoid having several agents repeat the same search.

## Keep one evidence ledger

Use stable IDs such as E01. Each row contains: claim, exact source URL or supplied file, location in source, date accessed, evidence type, audience/context, limitations and confidence. Keep quotations distinct from paraphrases. A forum comment is an individual signal, not a representative sample. Mark inaccessible or contradictory evidence explicitly.

## Synthesize after collection

Group observations by user task or friction, link every finding to evidence IDs and separate observation, interpretation and recommendation. Do not invent interviews, participants, quotations or adoption metrics. For contradictory sources, preserve the disagreement and explain which context may account for it.

## Report toward a decision

Deliver the decision in view, 3–5 supported findings, competing explanations, design implications and the smallest useful next research step. Include the evidence ledger. Label recommendations that rely on assumptions. If the evidence cannot answer the question, explain what is missing rather than asserting a conclusion.

## Draft validation still pending

Evaluate consistency across repeated briefs, recovery from unavailable sources and calibration of confidence against a human-reviewed evidence set. The draft is not a claim of validated automated research quality.