Deep Dive
Investigate: $ARGUMENTS
Agent Strategy
Every deep dive uses two core agents in parallel, plus optional specialists based on the query.
Core Agents (always spawn)
| Agent | Role | Strategy |
|---|---|---|
| Scout | Breadth-first explorer | Maps file structure, identifies hot spots, builds topology |
| Tracer | Depth-first investigator | Picks entry points, follows call chains deep into the stack |
Optional Agents (spawn when relevant)
| Agent | When to spawn | Strategy |
|---|---|---|
| Archaeologist | Query involves "why", history, or decisions | Digs through git blame/log for intent behind code |
| Boundary Mapper | Query involves integrations, APIs, or "what connects to" | Maps module boundaries, API surfaces, integration seams |
References: breadth-first-agent, depth-first-agent, context-archaeologist, boundary-mapper
Decision: Which Agents to Spawn
Read the user's query and decide:
- Always spawn Scout + Tracer (2 agents minimum)
- Add Archaeologist if the query asks why something exists, when it changed, or what motivated a design
- Add Boundary Mapper if the query asks about connections, integration points, API surfaces, or module interactions
- Maximum 4 agents — never more
Execution
1. Analyze the Query
Before spawning agents, identify:
- Target area: What part of the codebase to explore
- Depth level: Quick overview vs. deep investigation
- Query type: "How does X work?" vs. "Why is X like this?" vs. "What connects to X?"
2. Spawn Agents in Parallel
Use Task tool with subagent_type=Explore for each agent. Give each agent:
- The user's query for context
- Their specific investigation strategy (from reference files)
- Concrete starting points if you can infer them from the query
3. Synthesize
After all agents return, combine their findings using the Investigation Summarizer approach.
References: investigation-summarizer, output-format
Output
Present findings in the layered format — TL;DR first, then progressively detailed sections. The user should get the gist from the first 3 sentences and can read deeper as needed.
References: output-format
Anti-Patterns
- Don't spawn 4 agents for a simple "where is X defined?" question — use Grep directly
- Don't let agents read the same files — Scout maps topology, Tracer goes deep; they cover different ground
- Don't skip the synthesis step — raw agent outputs are disjointed; the summary is the value
- Don't present findings without file:line references — every claim needs a source