Parallel Execution Patterns

Parallel Execution Patterns

When to Load

• Trigger: Multi-agent tasks, concurrent operations, spawning subagents, parallelizing independent work
• Skip: Single-step tasks or sequential workflows with no parallelization opportunity

Core Concept

Parallel execution spawns multiple subagents simultaneously using the Agent tool (named Task before Claude Code 2.1.63; Task still works as an alias). Subagents run in the background by default, so N tasks run concurrently, dramatically reducing total execution time.

Critical Rule: ALL Agent calls MUST be in a SINGLE assistant message for true parallelism. If the calls are in separate messages, they launch one after another.

Execution Protocol

Step 1: Identify Parallelizable Tasks

Before spawning, verify tasks are independent:

• No task depends on another's output
• Tasks target different files or concerns
• Can run simultaneously without conflicts

Step 2: Prepare Dynamic Subagent Prompts

Each subagent receives a custom prompt defining its role:

You are a [ROLE] specialist for this specific task.

Task: [CLEAR DESCRIPTION]

Context:
[RELEVANT CONTEXT ABOUT THE CODEBASE/PROJECT]

Files to work with:
[SPECIFIC FILES OR PATTERNS]

Output format:
[EXPECTED OUTPUT STRUCTURE]

Focus areas:
- [PRIORITY 1]
- [PRIORITY 2]

Step 3: Launch All Tasks in ONE Message

CRITICAL: Make ALL Agent calls in the SAME assistant message:

I'm launching N parallel subagents:

[Agent 1]
description: "Subagent A - [brief purpose]"
prompt: "[detailed instructions for subagent A]"

[Agent 2]
description: "Subagent B - [brief purpose]"
prompt: "[detailed instructions for subagent B]"

[Agent 3]
description: "Subagent C - [brief purpose]"
prompt: "[detailed instructions for subagent C]"

On Claude Code versions that still run subagents in the foreground by default, add run_in_background: true to each call.

Step 4: Collect Results

Each subagent returns its final result to the parent conversation automatically when it finishes. Wait until every subagent has reported before synthesizing; do not poll, and do not start dependent work early. (The separate TaskOutput call is deprecated.)

Step 5: Synthesize Results

Combine all subagent outputs into unified result:

• Merge related findings
• Resolve conflicts between recommendations
• Prioritize by severity/importance
• Create actionable summary

Dynamic Subagent Patterns

Pattern 1: Task-Based Parallelization

When you have N tasks to implement, spawn N subagents:

Plan:
1. Implement auth module
2. Create API endpoints
3. Add database schema
4. Write unit tests
5. Update documentation

Wave 1 - spawn 3 subagents (independent of each other):
- Subagent 1: Implements auth module
- Subagent 2: Creates API endpoints
- Subagent 3: Adds database schema

Wave 2 - after wave 1 has finished (these depend on its output):
- Subagent 4: Writes unit tests
- Subagent 5: Updates documentation

Pattern 2: Directory-Based Parallelization

Analyze multiple directories simultaneously:

Directories: src/auth, src/api, src/db

Spawn 3 subagents:
- Subagent 1: Analyzes src/auth
- Subagent 2: Analyzes src/api
- Subagent 3: Analyzes src/db

Pattern 3: Perspective-Based Parallelization

Review from multiple angles simultaneously:

Perspectives: Security, Performance, Testing, Architecture

Spawn 4 subagents:
- Subagent 1: Security review
- Subagent 2: Performance analysis
- Subagent 3: Test coverage review
- Subagent 4: Architecture assessment

Task List Integration

When using parallel execution, task tracking (TaskCreate/TaskUpdate, or TodoWrite on older versions) differs:

Sequential execution: Only ONE task in_progress at a time Parallel execution: MULTIPLE tasks can be in_progress simultaneously

# Before launching parallel tasks
todos = [
  { content: "Task A", status: "in_progress" },
  { content: "Task B", status: "in_progress" },
  { content: "Task C", status: "in_progress" },
  { content: "Synthesize results", status: "pending" }
]

# As each subagent reports back, mark its task completed
todos = [
  { content: "Task A", status: "completed" },
  { content: "Task B", status: "completed" },
  { content: "Task C", status: "completed" },
  { content: "Synthesize results", status: "in_progress" }
]

When to Use Parallel Execution

Good candidates:

• Multiple independent analyses (code review, security, tests)
• Multi-file processing where files are independent
• Exploratory tasks with different perspectives
• Verification tasks with different checks
• Feature implementation with independent components

Avoid parallelization when:

• Tasks have dependencies (Task B needs Task A's output)
• Sequential workflows are required (commit -> push -> PR)
• Tasks modify the same files (risk of conflicts)
• Order matters for correctness

Performance Benefits

Approach	5 Tasks @ 30s each	Total Time
Sequential	30s + 30s + 30s + 30s + 30s	~150s
Parallel	All 5 run simultaneously	~30s

Parallel execution is approximately Nx faster where N is the number of independent tasks.

Example: Feature Implementation

User request: "Implement user authentication with login, registration, and password reset"

Orchestrator creates plan:

1. Implement login endpoint
2. Implement registration endpoint
3. Implement password reset endpoint
4. Add authentication middleware
5. Write integration tests

Parallel execution:

Wave 1 - launching 4 subagents in parallel:

[Agent 1] Login endpoint implementation
[Agent 2] Registration endpoint implementation
[Agent 3] Password reset endpoint implementation
[Agent 4] Auth middleware implementation

[Results arrive as each subagent finishes]

Wave 2 - depends on wave 1:

[Agent 5] Integration test writing

[Synthesize into cohesive implementation]

Troubleshooting

Tasks running sequentially?

• Verify ALL Agent calls are in a SINGLE message
• On older Claude Code versions, check run_in_background: true is set for each

Results not available?

• Results are delivered when each subagent finishes; wait for all of them
• A subagent that was denied a permission may return without finishing its work; check its report

Conflicts in output?

• Ensure tasks don't modify same files
• Add conflict resolution in synthesis step