Parallel Exploration

Patterns for launching multiple subagents in parallel to scout codebase, verify implementation, and coordinate via native Tasks.

Parallel Exploration (Scouting)

Launch multiple Explore subagents simultaneously when needing to find:

Related files across different areas
Similar implementations/patterns
Dependencies and usage

Pattern:

Task(subagent_type="Explore", prompt="Find [X] in [area1]", description="Scout area1")
Task(subagent_type="Explore", prompt="Find [Y] in [area2]", description="Scout area2")
Task(subagent_type="Explore", prompt="Find [Z] in [area3]", description="Scout area3")

Example - Multi-area scouting:

// Launch in SINGLE message with multiple Task calls:
Task("Explore", "Find auth-related files in src/", "Scout auth")
Task("Explore", "Find API routes handling users", "Scout API")
Task("Explore", "Find test files for auth module", "Scout tests")

Parallel Verification (Bash)

Launch multiple Bash subagents to verify implementation from different angles.

Pattern:

Task(subagent_type="Bash", prompt="Run [command1]", description="Verify X")
Task(subagent_type="Bash", prompt="Run [command2]", description="Verify Y")

Example - Multi-verification:

// Launch in SINGLE message:
Task("Bash", "Run typecheck: bun run typecheck", "Verify types")
Task("Bash", "Run lint: bun run lint", "Verify lint")
Task("Bash", "Run build: bun run build", "Verify build")

Task-Coordinated Parallel (Moderate+)

For multi-phase fixes, use native Tasks to coordinate parallel agents. See references/task-orchestration.md for full patterns.

Pattern - Parallel issue trees:

// Create separate task trees per independent issue
T_A1 = TaskCreate(subject="[Issue A] Debug", activeForm="Debugging A")
T_A2 = TaskCreate(subject="[Issue A] Fix",   activeForm="Fixing A",   addBlockedBy=[T_A1])
T_B1 = TaskCreate(subject="[Issue B] Debug", activeForm="Debugging B")
T_B2 = TaskCreate(subject="[Issue B] Fix",   activeForm="Fixing B",   addBlockedBy=[T_B1])
T_final = TaskCreate(subject="Integration verify", addBlockedBy=[T_A2, T_B2])

// Spawn agents per issue tree
Task("fullstack-developer", "Fix Issue A. Claim tasks via TaskUpdate.", "Fix A")
Task("fullstack-developer", "Fix Issue B. Claim tasks via TaskUpdate.", "Fix B")

Agents claim work via TaskUpdate(status="in_progress") and complete via TaskUpdate(status="completed"). Blocked tasks auto-unblock when dependencies resolve.

When to Use Parallel

Scenario	Parallel Strategy
Root cause unclear, multiple suspects	2-3 Explore agents on different areas
Multi-module fix	Explore each module in parallel
After implementation	Bash agents for typecheck + lint + build
Before commit	Bash agents for test + build + lint
2+ independent issues	Task trees per issue + fullstack-developer agents

Combining Explore + Tasks + Bash

Step 1: Parallel Explore to scout Step 2: Sequential implementation (update Tasks as phases complete) Step 3: Parallel Bash to verify

// Scout phase - parallel
Task("Explore", "Find payment handlers", "Scout payments")
Task("Explore", "Find order processors", "Scout orders")

// Wait for results, implement fix, TaskUpdate each phase

// Verify phase - parallel
Task("Bash", "Run tests: bun test", "Run tests")
Task("Bash", "Run typecheck", "Check types")
Task("Bash", "Run build", "Verify build")

Resource Limits

Max 3 parallel agents recommended (system resources)
Each subagent has 200K token context limit
Keep prompts concise to avoid context bloat
Use TaskList() to check for available unblocked work

3.6 KiB Raw Blame History

Parallel Exploration