Multi-Agent Systems
Multi-agent systems split work across specialized agents.
This is useful only when specialization improves control. More agents do not automatically mean better results.
Common patterns
| Pattern | Use case |
|---|---|
| supervisor | one planner delegates work |
| sequential pipeline | each agent transforms the artifact |
| parallel reviewers | independent checks reduce blind spots |
| debate | explore competing answers |
| specialist tools | each agent owns a tool domain |
Shared state
Agents should work from shared task state:
- objective
- plan
- evidence
- open questions
- completed work
- final artifact
Avoid hidden side conversations that cannot be audited.
Risks
- agents repeat each other
- cost explodes
- responsibility becomes unclear
- one bad agent poisons the rest
- unsafe delegation bypasses approval
Design rule
Start with one agent. Add another only when it has a clear job, separate tools, and measurable value.
Knowledge check
Q1: When should you add another agent?
When a distinct role improves quality, safety, or parallelism enough to justify complexity.
Q2: Why does shared state matter?
It makes progress auditable and prevents agents from losing context.