Deep Research

2. Spawn children with focused task descriptions

Each child gets a clear, self-contained task. Include:

• What to research
• Where to look (web search, GitHub, specific docs, specific repos)
• What to return — the format you expect (bullet points, comparison table, pros/cons)

Use descriptive labels: research-columnar-dbs, audit-rest-endpoints, read-nextauth-docs.

3. Wait for all completions

Do NOT poll. Track expected child IDs and wait for all TaskCompletionEvent messages.

4. Synthesize

Once all children report back:

• Cross-reference findings — look for agreement and contradiction between children
• Identify gaps — if an angle was missed or a child came back thin, spawn a follow-up child
• Rank and recommend — don't just dump information; give the user your assessment
• Cite sources — attribute findings to specific children or sources so the user can dig deeper

5. Present as a structured briefing

Format:

## Research: {topic}

### Key Findings
- ...

### Comparison (if applicable)
| Option | Pros | Cons | Verdict |

### Recommendation
...

### Open Questions
- Things that need further investigation or user input

Important Rules

• Always spawn at least 2 children — single-threaded research defeats the purpose
• Never present raw child output — always synthesize and add your own analysis
• Spawn follow-up children if initial results are incomplete or contradictory
• Be honest about confidence — flag when findings are thin or sources disagree
• Prefer primary sources (official docs, source code, APIs) over blog posts and opinions
• Children should use bash_exec for web searches, reading repos, and fetching docs