Parallel Execution

Execute multiple independent tasks simultaneously to maximize throughput and minimize total execution time.

When to Use

• Multiple independent tasks (no dependencies)
• Tasks benefit from concurrent execution
• Maximizing throughput is priority
• Available agents for parallel work
• Results can be aggregated after completion

Core Concepts

Independence

Tasks are independent when:

• ✓ No data dependencies (one doesn’t need other’s output)
• ✓ No resource conflicts (different files, databases)
• ✓ No ordering requirements (either can complete first)
• ✓ Failures are isolated (one failing doesn’t block others)

Example – Independent:

✓ Task A: Review code quality (code-reviewer)
✓ Task B: Run test suite (test-runner)
→ Can run in parallel

Example – NOT Independent:

✗ Task A: Implement feature (feature-implementer)
✗ Task B: Test feature (test-runner)
→ B depends on A's output, must run sequentially

Concurrency

Critical: Use single message with multiple Task tool calls

Correct:

Send one message containing:
- Task tool call #1 → Agent A
- Task tool call #2 → Agent B
- Task tool call #3 → Agent C

All three agents start simultaneously.

Incorrect:

Message 1: Task tool → code-reviewer
[wait]
Message 2: Task tool → test-runner
[wait]

This is sequential, NOT parallel!

Synchronization

Collection Point:

• Wait for all parallel agents to complete
• Collect results from each agent
• Validate each result independently
• Aggregate results into final output

Parallel Execution Process

Step 1: Identify Independent Tasks

Independence Checklist:

• No data dependencies
• No shared writes (read-only or different targets)
• No execution order requirements
• Failures don’t cascade
• Results can be validated independently

Step 2: Agent Assignment

Match Tasks to Agents:

Task 1: Review code quality → code-reviewer
Task 2: Run test suite → test-runner
Task 3: Run benchmarks → test-runner

Agent Availability Check:

• Ensure sufficient agents available
• Check for specialization overlap
• Consider workload balance

Step 3: Launch Parallel Execution

Single message with multiple Task tool calls:

<Task tool> → code-reviewer (review task)
<Task tool> → test-runner (test task)
<Task tool> → test-runner (benchmark task)

All agents start simultaneously.

Step 4: Monitor Execution

Track Progress:

Agent 1 (code-reviewer): In Progress
  Task: Code quality review

Agent 2 (test-runner): Completed ✓
  Task: Test suite
  Result: 45/45 tests passed

Agent 3 (test-runner): In Progress
  Task: Benchmarks

Step 5: Collect & Validate Results

As Each Agent Completes:

1. Collect output
2. Validate against success criteria
3. Check for errors
4. Mark as complete or failed

Step 6: Aggregate Results

## Parallel Execution Results

### Completed Tasks:
1. ✓ Code quality review (code-reviewer)
   - Result: 3 minor issues found

2. ✓ Test suite (test-runner)
   - Result: All tests passing (45/45)

3. ✓ Performance benchmarks (test-runner)
   - Result: All benchmarks acceptable

### Overall Status: ✓ Success (with minor issues)

Execution Patterns

Pattern 1: Homogeneous Parallel

All agents same type, different inputs:

Use Case: Test multiple modules

├─ test-runner: Test memory-core
├─ test-runner: Test memory-storage-turso
└─ test-runner: Test memory-storage-redb

Single message, 3 Task tool calls

Pattern 2: Heterogeneous Parallel

Different agent types, related task:

Use Case: Comprehensive code check

├─ code-reviewer: Quality analysis
├─ test-runner: Test execution
└─ debugger: Performance profiling

Single message, 3 Task tool calls

Pattern 3: Parallel with Convergence

Parallel execution → Single synthesis:

Phase 1: Parallel Investigation
├─ debugger: Profile performance
├─ code-reviewer: Analyze efficiency
└─ test-runner: Run benchmarks

[All complete]

Phase 2: Synthesis
└─ Combine findings, identify root cause

Synchronization Strategies

Wait for All (AND)

Most Common:

• Wait for ALL agents to complete
• Proceed only when all finished
• Useful when all results needed

Wait for Any (OR)

Early Termination:

• Proceed when ANY agent completes successfully
• Cancel or continue others
• Useful for redundant approaches

Wait for Threshold

Partial Completion:

• Proceed when N out of M agents complete
• Useful for resilience
• Handle missing results gracefully

Resource Management

Available Agents

• code-reviewer (1 instance)
• test-runner (1 instance)
• feature-implementer (1 instance)
• refactorer (1 instance)
• debugger (1 instance)

Parallelization Limit: Maximum 5 agents simultaneously (one of each type)

Workload Balancing

Distribute Evenly:

Tasks: [T1, T2, T3, T4, T5, T6]
Agents: [A, B, C]

Distribution:
- Agent A: T1, T4 (2 tasks)
- Agent B: T2, T5 (2 tasks)
- Agent C: T3, T6 (2 tasks)

Error Handling

Independent Failures

Isolation:

• One agent failing doesn’t stop others
• Continue collecting successful results
• Report failed tasks separately

Parallel Execution:
├─ Agent A: ✓ Success
├─ Agent B: ✗ Failed (error in code)
└─ Agent C: ✓ Success

Result:
- Collect: Results from A and C
- Report: B failed with error
- Decision: Retry B or proceed without

Partial Success Handling

Strategies:

1. Fail Fast: If any fails, stop and report
2. Best Effort: Collect all successful results
3. Retry Failed: Let successful complete, retry failures

Performance Optimization

Speedup Calculation

Sequential time = T1 + T2 + T3 + ... + Tn
Parallel time = max(T1, T2, T3, ..., Tn)

Speedup = Sequential time / Parallel time

Example:

Tasks:
- Task A: 10 minutes
- Task B: 15 minutes
- Task C: 8 minutes

Sequential: 10 + 15 + 8 = 33 minutes
Parallel: max(10, 15, 8) = 15 minutes

Speedup: 33 / 15 = 2.2x faster

Optimal Parallelization

Identify Bottlenecks:

• Find longest-running task
• Can it be decomposed further?
• Can it be optimized?
• Start slow tasks first

Best Practices

DO:

✓ Verify independence before parallelizing ✓ Use single message with multiple Task tool calls ✓ Balance workload across agents ✓ Set appropriate timeouts for each task ✓ Handle failures gracefully (isolation) ✓ Validate each result independently ✓ Aggregate comprehensively at the end

DON’T:

✗ Parallelize dependent tasks ✗ Send sequential messages thinking they’re parallel ✗ Overload single agent while others idle ✗ Skip validation because “it’s parallel” ✗ Assume all will succeed ✗ Ignore agent failures

Examples

Example 1: Simple Parallel Review

Task: "Check code quality and run tests"

Analysis: Independent tasks, different agents

Plan:
├─ code-reviewer: Review code quality
└─ test-runner: Run test suite

Execution: [Single message with 2 Task tool calls]

Results:
- code-reviewer: 2 issues found ✓
- test-runner: 45/45 tests pass ✓

Speedup: 2x

Example 2: Multi-Module Testing

Task: "Test all crates in workspace"

Analysis: 3 independent crates, same agent type

Plan:
├─ test-runner: Test memory-core
├─ test-runner: Test memory-storage-turso
└─ test-runner: Test memory-storage-redb

Execution: [Single message with 3 Task tool calls]

Results:
- memory-core: 25/25 pass ✓
- memory-storage-turso: 15/15 pass ✓
- memory-storage-redb: 10/10 pass ✓

Speedup: 3x

Example 3: Comprehensive Quality Check

Task: "Pre-release quality validation"

Analysis: 4 independent checks, maximum parallelization

Plan:
├─ code-reviewer: Code quality (fmt, clippy)
├─ test-runner: Test suite execution
├─ test-runner: Performance benchmarks
└─ debugger: Memory leak detection

Execution: [Single message with 4 Task tool calls]

Results: All checks pass ✓

Speedup: 4x

Integration

Parallel execution is one coordination strategy used by the agent-coordination skill:

Coordination Strategy Selection:
├─ Independent tasks → Use parallel-execution (this skill)
├─ Dependent tasks → Use sequential coordination
├─ Complex mix → Use hybrid coordination
└─ Multiple perspectives → Use swarm (with parallel)

Summary

Parallel execution maximizes efficiency for independent tasks by:

• Concurrent agent execution (single message, multiple tools)
• Independent task validation (no cross-dependencies)
• Synchronized result collection (wait for completion)
• Comprehensive aggregation (synthesize final output)

When done correctly, parallel execution provides significant speedup while maintaining quality and reliability.