Exploring in Parallel
Overview
Parallelize exploration aggressively. When researching or exploring, launch multiple subagents simultaneously instead of sequential searches. This reduces wall-clock time proportionally to the number of parallel agents.
How to Parallelize
Use the Task tool with subagent_type=Explore. The key is to send multiple Task calls in a single message — this launches them concurrently.
Each subagent gets:
subagent_type:"Explore"for codebase searches,"general-purpose"for web researchdescription: Short label (3-5 words)prompt: Detailed, self-contained search instructions
Example — investigating how auth works:
Send in ONE message:
Task 1: { subagent_type: "Explore", description: "Find auth middleware", prompt: "Search for authentication middleware, auth guards, session validation..." }
Task 2: { subagent_type: "Explore", description: "Find login endpoints", prompt: "Search for login routes, /auth endpoints, credential handling..." }
Task 3: { subagent_type: "Explore", description: "Find auth config", prompt: "Search for JWT config, OAuth settings, auth environment variables..." }
All three run simultaneously. When results return, synthesize findings into a coherent answer.
When to Parallelize
| Situation | Action | |-----------|--------| | Searching for multiple patterns | One subagent per pattern | | Exploring different directories | One subagent per area | | Investigating related questions | One subagent per question | | Checking multiple files (>3) | Parallel reads | | Web research on multiple topics | One subagent per topic |
When NOT to Parallelize
- Dependent searches: Second search needs results from first
- Single specific file: Just use
Readdirectly - Simple grep: One pattern, one location — use
Grepdirectly - Known target: Know exact file/function — use
Glob/Readdirectly
Red Flags — STOP and Parallelize
| Thought | Action | |---------|--------| | "Let me search for X first, then Y" | Launch both simultaneously | | "I'll check this file, then that one" | Parallel reads if >3 files | | "First explore area A, then B" | Parallel subagents per area | | "Let me see what this returns before..." | Usually can parallelize anyway |
After Results Return
Don't just dump raw results. Synthesize:
- Identify overlapping findings across agents
- Resolve contradictions
- Present a unified answer with file references
- Note any gaps that need follow-up
Scan to join WeChat group