What is agent choreography?
Updated November 2026
Agent choreography is a coordination style in which agents react to events or shared signals rather than receiving every instruction from a central controller. Each participant follows local rules, and their interactions produce the overall workflow.
How does choreography differ from orchestration?
Microsoft’s agent architecture guide documents explicit coordination patterns such as sequential execution and handoffs. Choreography describes a different control arrangement: participants decide locally when to act. It is an architectural choice, not a synonym for any multi-agent conversation.
For example, a research agent publishes a completed evidence set, a checking agent reacts to that event, and a drafting agent proceeds only after verification. No single agent needs to issue every instruction, but the event contracts still define the permitted progression.
What needs to be in an event contract?
Include an event identifier, task correlation identifier, event type, schema version, producing agent, and references to relevant artifacts. Specify whether an event reports a completed fact or requests future work. Confusing these meanings can trigger actions before their prerequisites are satisfied.
| Concern | Design rule |
|---|---|
| Duplicate delivery | Track handled events or make actions idempotent |
| Out-of-order arrival | Check prerequisites before acting |
| Ownership | Name the agent responsible for each action |
Where does decentralized coordination break down?
Local decisions can create global loops. One agent’s update triggers another update that triggers the first agent again. Bound reaction chains, distinguish new information from repeated observations, and define terminal task states that suppress further work.
Choreography reduces dependence on a central decision point but makes end-to-end behavior harder to inspect. Use correlated event histories and explicit ownership rules so operators can answer why an action occurred and which participant should recover a failure.
Do not assume that an event bus supplies exactly-once business execution. Delivery behavior depends on the chosen infrastructure, and retries can still duplicate external effects. For workflows with strict ordering or many compensating actions, a central orchestrator may be easier to reason about than a growing network of reactions.
What else should you know?
Does choreography mean no rules?
No. It relies on clear event contracts and local transition rules.
Is a shared chat room sufficient?
No. Visibility alone does not define triggers or execution ownership.
Can orchestration and choreography coexist?
Yes. Local workflows can be orchestrated inside a wider event-driven system.
What should be tested first?
Test duplicates, delayed events, circular triggers, and unavailable participants.