Create Agent

An agent is a harness, a model, and its own directories, hosted by one supervisor.
Back to Agents
What creating this agent does.
The supervisor writes a durable agent record and creates the agent's directories. No process starts and no quota is consumed.
The agent starts signed out. Its configuration directory is created empty, so the first conversation you open with it has to sign the harness CLI in once. Every later conversation of this agent reuses that sign-in.
Each conversation keeps its own workspace. A conversation gets its own copy of the agent's initial working directory, and the copy is kept after the conversation ends because it holds that conversation's work. Nothing reclaims them yet, so an agent's disk use on its host only grows.
Tool permissions stay fleet-wide for now: an agent's own tool bindings are stored but conversations still materialize the fleet-global tool configuration, so there is nothing per-agent to choose here yet.
Cancel