Deep research

Deep Research (Agent Teams)

Conduct deep, multi-source research using a coordinated Agent Team. Multiple researcher agents work in parallel on different sub-topics, share discoveries via messaging, and the team lead synthesizes findings into a structured report.

Architecture

Team Lead (you)
├── researcher-1  (sub-topic A)
├── researcher-2  (sub-topic B)
├── researcher-3  (sub-topic C)
└── ...up to researcher-6

• Team Lead: Decomposes topic, creates tasks, spawns researchers, synthesizes final report.
• Researchers: general-purpose agents with WebSearch/WebFetch. Each owns one sub-topic task. Reports findings back via SendMessage.

Workflow

Step 1: Decompose the Research Topic

Break the user's topic into 3-6 independent sub-questions. Select from these angles based on relevance:

Angle	Example sub-question
Background	What is X? History and origin.
Current State	Latest developments, versions, adoption.
Comparisons	Alternatives, trade-offs, benchmarks.
Best Practices	Recommended approaches, common pitfalls.
Ecosystem	Community, tooling, integrations.
Future	Roadmap, predictions, emerging trends.

If the topic is too broad, ask the user to narrow it first.

Step 2: Create the Research Team

TeamCreate:
  team_name: "deep-research-{topic-slug}"
  description: "Deep research on {topic}"
  agent_type: "research-lead"

Step 3: Create Tasks for Each Sub-question

Use TaskCreate for each sub-question:

TaskCreate:
  subject: "Research: {sub-question title}"
  description: "{detailed sub-question with scope}"
  activeForm: "Researching {sub-question title}"

Step 4: Spawn Researcher Agents in Parallel

Launch all researchers in a single message using the Task tool. Each researcher:

• Has subagent_type: "general-purpose"
• Joins the team via team_name
• Gets a unique name (e.g., researcher-1)
• Runs in background via run_in_background: true

Researcher prompt template:

You are a research agent on team "{team-name}". Your name is "{researcher-name}".

## Your Task
Research this question thoroughly: {sub-question}

## Instructions
1. Read the team config at ~/.claude/teams/{team-name}/config.json to discover teammates
2. Claim your task using TaskUpdate (set owner to your name, status to in_progress)
3. Use WebSearch to search at least 3 different queries related to your question
4. Use WebFetch to read the top 2-3 most relevant results
5. Extract key findings with specific facts, numbers, and quotes
6. Note any conflicting information across sources
7. Mark your task as completed using TaskUpdate
8. Send your findings to the team lead using SendMessage

## Output Format
Send findings via SendMessage to the team lead with this structure:

### {Sub-question Title}

**Key Findings:**
- [Finding 1 with specific data]
- [Finding 2 with specific data]
- ...

**Conflicts/Uncertainties:**
- [Any disagreements between sources]

**Sources:**
- [Title](URL) - What this source contributed

Step 5: Monitor and Synthesize

As researchers send back findings via SendMessage:

1. Acknowledge receipt (optional — only if a researcher needs follow-up)
2. After all researchers report back, synthesize into the final report
3. Shut down all researchers via SendMessage with type: "shutdown_request"

Step 6: Present the Research Report

Format:

# Research Report: [Topic]

## Executive Summary
[2-3 paragraph overview of key findings across all sub-topics]

## Detailed Findings

### [Sub-topic 1]
[Synthesized findings]

### [Sub-topic 2]
[Synthesized findings]

...

## Key Takeaways
- [Most important conclusions]

## Conflicts & Open Questions
- [Areas where sources disagreed or data was insufficient]

## Sources
- [Deduplicated list of all sources with URLs]

Offer to dive deeper into any specific sub-topic.

Guidelines

• Parallel launch is critical: Spawn all researchers in a single message.
• Use background mode: Set run_in_background: true so researchers work concurrently.
• Source quality: Prefer official docs, reputable blogs, academic papers, established outlets.
• Conflict resolution: Present both perspectives when sources disagree.
• Recency: Prioritize recent sources; note publication dates.
• Language: Follow CLAUDE.md language rules for the report.
• Cleanup: Always shut down teammates when research is complete.