Multi-Agent Patterns
Outcome: You'll understand parallel processing with sub-agents and know when to use this pattern.
The Architecture
When a task is too big for one 200K window, you can parallelize.
How it works:
- You have a "master" agent (your main Claude session)
- Master spawns "sub-agents" for subtasks
- Each sub-agent gets its own fresh 200K tokens
- They work simultaneously
- Results merge back to master
The constraint: Communication between agents must be minimal. Each "report back" costs tokens. So sub-agents report simple signals: "done" or a brief summary.
When to Use Multi-Agent
Good for:
- Batch processing (enrich 50 documents)
- Independent research tasks (research 5 competitors separately)
- Large content generation (write 10 chapters)
Not good for:
- Tasks with sequential dependencies (step 2 needs step 1's output)
- Highly interconnected work
For sequential tasks: Use the handoff pattern with working logs instead.
Hands-On: Spawn a Sub-Agent
- Give Claude a task that has clear subtasks:
"I need you to research three topics: [Topic A], [Topic B], and [Topic C]. For each, create a summary with key findings. Spawn sub-agents to handle each topic in parallel."
- Watch Claude create the sub-agents
- See how it coordinates and merges results
Look for "Task" (that is a sign it's using a subagent)
Real Example
The setup: 68 playbooks needed enrichment with additional research.
The approach:
- Master agent orchestrated the work
- Spawned 5 sub-agents at a time
- Each sub-agent handled 1 playbook
- Sub-agents reported "done" when finished
- Master moved to next batch
The result: 13.8 million tokens used efficiently across all sub-agents. Human review only needed at the end.
Key Takeaways
| Pattern | Best For | Token Strategy |
|---|---|---|
| Single Agent | Focused tasks, sequential work | Use handoffs when approaching limit |
| Multi-Agent | Batch processing, parallel research | Each agent gets fresh 200K window |
| Hybrid | Complex projects | Master coordinates, spawns agents for independent subtasks |
The power of multi-agent patterns is turning a token constraint into an architecture advantage—you get parallel processing AND fresh context for each subtask.