A reasoning unit, with a job.
Name, type, behaviour prompt, and a toolbelt. The orchestrator can call any agent you save here.
Identity
Leave as default unless this agent needs something specific.
Behaviour prompt
Define what this agent does, step by step. Imperative voice; numbered steps where useful.
Toolbelt
Pick the tools this agent can call. The orchestrator will only expose these to the agent at runtime.
Sub-agents (optional)
For managers: agents this one can delegate to. Each becomes a callable delegate_to_<name> at runtime.