Multi-agent systems split work across several agents. They can help, but they add complexity.
Common Patterns
- Orchestrator-worker: a lead agent assigns sub-tasks to workers and combines results.
- Pipeline: agents handle successive stages — research, draft, review.
- Critic or reviewer: one agent checks another's work.
- Parallel exploration: several agents try different approaches; the best result wins.
When They Help
- Broad tasks with independent parts, such as research across many sources.
- Tasks exceeding a single context window.
- Workflows benefiting from independent review.
When They Don't
- Tightly coupled tasks where agents need constant shared context — coordination costs outweigh gains.
- Simple tasks a single agent handles well.
Challenges
- Higher token costs.
- Information lost between agents.
- Harder debugging and evaluation.
- Duplicated or conflicting work.
Advice
Start with one agent. Add more when evaluation shows a clear benefit, and give each agent a precise, self-contained task and output format.