Agents

An agent is an autonomous member of your organization: a principal with a role (its permissions), like a teammate but automated. It isn't a prompt or a model — it's an identity that does work.

What you create

Creating an agent makes a record (name + description) plus an org membership carrying a role. The role is what the agent is allowed to do.

• Create — organization_agents_create (name, description?, roleId)
• List — organization_agents_list

Any org member or admin can create agents — no platform admin.

Two ways an agent runs

Same agent identity, two execution paths:

          ┌─ connected: your Claude Code authenticates AS the agent (its token)
agent ────┤
          └─ managed:   a runner daemon claims & runs it for you

• Connected — point an existing agent (e.g. Claude Code) at BitsWeave with a token scoped to this agent; it acts as the agent. Same wiring as Getting Started, using the agent's token.
• Managed — a runner daemon registers, claims the agent's work, and runs it (operator-deployed).

Where its work comes from

An agent doesn't poll the world — work is routed to it.

broker event ──► routing rule ──► agent inbox ──► agent acts

A routing rule matches incoming broker events and drops them into the agent's inbox; the agent (connected or managed) claims them and acts.

Next: Create and route an agent.