Multi-agent systems
Several agents, one task
A multi-agent system splits work across agents with different roles, instructions, and tools. One might research, another write, another review.
It combines the earlier patterns: an orchestrator assigns the work, workers act in their own loops, and an evaluator checks the result.
1. Give each agent a clear role
- Lead agentSplits the brief into questions
- Research agentsEach one searches one question
- Writer agentDrafts from the findings
- Reviewer agentChecks sources and gaps
Separate contexts are the main benefit: each agent sees only what it needs, so a long research trail does not crowd the writer’s window.
2. Design the handoffs
Agents pass work through structured messages. Define what each handoff must contain, as in the structured outputs lesson, so nothing important is lost between agents.
3. Count the cost
Every agent adds model calls, latency, and new ways to fail. Mistakes can also spread: one agent’s wrong finding becomes another agent’s input.
4. Start with one agent
Many tasks that seem to need several agents work well with one agent and good tools. Split only when one context becomes too crowded or the roles truly need different tools and permissions.
When is it useful?
For broad tasks with separable parts, such as research across many sources, where parallel work and separate contexts clearly help. For most tasks, a single agent or a workflow is simpler, cheaper, and easier to debug.