Supervisors

Creating Supervisors

Define domain supervisors that own execution, evidence, and result quality.

Create supervisors around one domain and make their return contract stricter than their personality.

Choose a Domain

Start with a domain you expect to route often. Examples include release analysis, trace review, research summarization, or desktop execution. Name the supervisor after the domain, not after a personality.

Assign Agents

Attach only the agents that make sense inside that domain. If one supervisor needs every agent in the system, the domain boundary is probably wrong.

Set Tool Policy

The domain should determine which tools are available and what evidence they must produce. This is where safety levels and intent checks matter most.

Return Shape

Treat the output as a contract: completed, unverified, or failed. A supervisor that cannot distinguish those states will make every workflow look healthier than it is.

Next Steps

Continue with Understanding Traces so you know how to inspect the supervisor once it starts running real work.