| name | subagent-research |
| description | Offload verbose research (web, code, files, GitHub, video) to a subagent instead of burning main context window. |
Subagent Research
Golden rule: if the answer could be longer than 3 lines, use a subagent.
Core Pattern
- Identify the research need
- Pick the right tool (see table below)
- Craft 1-4 specific questions
- Dispatch a
default subagent with questions + brevity instruction
- Subagent runs research in its own context, returns terse answers
Template
Answer these questions as concisely as possible (one sentence per question):
1. [precise question]
2. [precise question]
No summarisation needed. Use [tool].
When NOT to Use
- Single fact — "What year was Python 3.12?" → tool directly
- Already know the query, answer is short → inline
- Need answer to proceed → no subagent overhead
- Trivial file lookup —
grep inline
Research Types
| Type | Tool | Example question |
|---|
| Web / general | web_search | "What are the pros/cons of X vs Y?" |
| Code / library internals | code_search | "What's the signature of React.useCallback?" |
| GitHub repos | fetch_content | "What does the README say about installation?" |
| YouTube / video | fetch_content + prompt | Pass user's exact question as prompt |
| File / workspace | bash (grep/find), read | "Where is DB configured in this project?" |
| Complex / multi-tool | combination | "Look up lib docs + check usage in codebase" |
Common Mistakes
- Vague questions → rambling answers. Be specific.
- No brevity instruction → subagent defaults to verbose synthesis
- 5+ questions → split into multiple subagent calls
- Wrong tool for the type →
web_search for code questions misses signatures
web_search on main agent → do it in a subagent instead