One host for every agent
Agents share one host. A new agent is a published spec, and the host runs it with the others.
Host for scheduled AI agents
Watchrun hosts scheduled AI agents. One host runs each agent on a schedule or from a hook, then records the run. Each agent follows a spec published in UniFact.
Model use for each agent is collected in AI Usage Metrics, in the same report as an employee’s coding agent. The row uses the agent name from its spec.
A schedule or a hook starts the run
An agent can run each weekday, or when a hook calls the host. The log records the spec, what it watched, the decision, and the action.