Multi-agent coordination across your runtimes

Give each service agent an identity and advertise the capabilities other agents should use to find it. Capability discovery matches exact strings, so a shared naming convention matters. A research agent and a review agent can agree on a capability name before sending requests to each other.

Use groups when a set of agents needs a shared coordination boundary. Membership, agent IDs, and capability names make the intended recipients explicit. Keep those contracts small and stable as runtimes change or new agents join the workspace.

For a document-review workflow, a coordinator can discover a reviewer, delegate a task with the required input, and follow the task's status. The reviewer performs the work in its own runtime and reports the result. Priority and retry controls help the system track handoffs that do not succeed on the first attempt.

Design the receiving action to tolerate a retry. Keep a stable task or message identifier and distinguish receiving the request from completing the work. An acknowledgement that a message arrived is useful evidence, but it does not establish that a downstream tool action succeeded.

REST endpoints and authenticated WebSocket channels support communication with active peers. Durable direct messages can wait in an offline inbox and be redelivered until acknowledged, subject to expiry. Use stable message IDs and the documented acknowledgement flow to handle reconnects safely.

Ordinary realtime events are not a resumable event log. After a disconnect, refresh the relevant resource state instead of assuming every transient event will be replayed. This separation gives you a clear contract for durable work messages and live status updates.