Creativity · Agent Protocol
Orchestrator-Worker Pattern
Orchestrator-worker is the most common multi-agent pattern in production. A lead 'orchestrator' agent plans the task, decides which sub-task goes to which specialised worker, aggregates the returned results, and decides whether to continue or stop. Workers are typically narrow, well-tested agents with small tool surfaces. Anthropic's own 'Building Effective Agents' post describes this as a core pattern for real-world systems.
Protocol facts
- Sponsor
- open pattern (Anthropic, OpenAI, Google)
- Status
- stable
- Spec
- https://www.anthropic.com/research/building-effective-agents
- Interop with
- A2A, LangGraph, CrewAI, AutoGen, Claude subagents
Frequently asked questions
When should I use orchestrator-worker?
When the task naturally decomposes into sub-tasks of different shapes — research + writing + verification, or triage + routing + resolution. The orchestrator keeps the plan; workers don't need to know about each other.
How big should workers be?
Small and narrow. Each worker ideally has one job, a small set of tools, and a prompt short enough to reason about. That's what makes orchestrator-worker robust — failures are localised.
Isn't this just microservices?
In spirit, yes — which is why it feels natural to experienced distributed-systems engineers. The difference is that the orchestrator uses an LLM to plan rather than a hardcoded workflow.
Sources
- Anthropic — Building Effective Agents — accessed 2026-04-20