parallel-execution

name	parallel-execution
description	Parallel execution pattern: Multiple agents work simultaneously on independent sub-tasks.

---

name: "parallel-execution" description: "Parallel execution pattern: Multiple agents work simultaneously on independent sub-tasks. Best for embarrassingly parallel workloads."

Parallel Execution Pattern

[!IMPORTANT] When to use: When sub-tasks are independent and can be executed simultaneously. Maximizes throughput for parallelizable workloads.

Decision Tree Path

Does the task require multiple specialized capabilities?
├── YES
│   └── Are subtasks independent (no shared state/collaboration)?
│       ├── YES → Parallel Execution (you're here)
│       └── NO, need shared context?
│           └── YES → supervisor-orchestrator
└── NO → single-agent

When to Choose

✅ Use Parallel When	❌ Avoid When
Subtasks are truly independent	Subtasks share state/artifacts
You need maximum throughput	Order dependency exists
Each subtask is self-contained	Errors in one affect others
Subtasks are embarrassingly parallel	Result synthesis is complex

Use Cases

• Batch document processing (N docs, each processed independently)
• A/B content generation (multiple variants simultaneously)
• Multi-source research (search multiple sources at once)
• Grid of evaluations (test N configurations in parallel)
• User research synthesis (analyze N surveys simultaneously)

Example Architecture

┌─────────────────────────────────────────────────────┐
│                    Dispatcher                       │
│         (Splits task, aggregates results)            │
└────────────────────┬───────────────────────────────┘
                     │
    ┌────────────────┼────────────────┐
    ▼                ▼                ▼
┌────────┐      ┌────────┐      ┌────────┐
│ Agent 1│      │ Agent 2│      │ Agent 3│  ← All simultaneous
│(Doc A) │      │(Doc B) │      │(Doc C) │
└────────┘      └────────┘      └────────┘
    │                │                │
    └────────────────┼────────────────┘
                     ▼
              ┌─────────────┐
              │  Aggregator  │
              │ (Synthesize) │
              └─────────────┘
                     │
                     ▼
                 Final Output

Dispatcher Pattern

You are a dispatcher for parallel execution.

Task: {user_task}

1. Identify N independent subtasks
2. Launch all subtasks simultaneously
3. Wait for all completions
4. Aggregate results

Subtasks identified: {list}

Key Considerations

Include in Your Decision

1. True independence — Can any subtask fail without affecting others?
2. Result aggregation — How will results be combined?
3. Load balancing — Are subtasks similar in complexity?
4. Error handling — Partial failure vs. total failure?

Independence Checklist

Question	If NO, Use Another Pattern
Do subtasks need shared state?	❌ → supervisor-orchestrator
Do subtasks depend on each other's output?	❌ → supervisor-orchestrator
Is there an execution order requirement?	❌ → supervisor-orchestrator
Do subtasks need to collaborate?	❌ → supervisor-orchestrator

Anti-Patterns

Mistake	Why It's Bad
Faking independence	Race conditions, corrupted outputs
No aggregation plan	Results pile up without synthesis
Uneven workload	Some agents idle, others overwhelmed
Ignoring partial failures	Incomplete results, wasted compute

Aggregation Strategies

Strategy	Use When
Simple concatenation	Results are independent, list format OK
Weighted scoring	Some results more important than others
Consensus voting	Multiple perspectives on same question
Hierarchical synthesis	Results themselves need synthesis

Scaling Up: Adding Dependencies

Need some coordination between parallel tasks?
├── YES, light handoffs?
│   └── YES → supervisor-orchestrator (light version)
└── YES, hierarchical control?
    └── YES → hierarchical-supervisor

Scaling Down

Is parallelism actually overkill?
├── YES, one LLM can handle it?
│   └── YES → single-agent
└── YES, simple routing by type?
    └── YES → supervisor-delegation

Quick Reference

Factor	Parallel Execution
Latency	Low (limited by slowest subtask)
Cost	Higher (N agents running)
Complexity	Low-Moderate
Reliability	Partial failure possible; others continue
Best for	Independent batch workloads

---

Pattern ID: parallel-execution | AGEM-compatible