Agents
Compose model and Tool execution with explicit run state, context and application ownership.
Run application work, not your coding assistant
An Agent combines a definition with an execution runtime. The definition declares identity, instructions and available Tools; the runtime supplies Worker execution, inference and governance bindings.
Use an embedded runtime for single-process execution, or borrow a host-owned Worker when it needs an independent lifetime. Neither creates a universal application instance.
- Define and run an agent for the first executable path.
- Instructions and model context for trusted policy and bounded history.
- Sessions and conversation threads for distinct persistence responsibilities.
- Application execution ownership for admission and joined cleanup.
- Streaming, cancellation and recovery for observable runs.
Add only the capabilities you need
Models can run without an Agent. Tools can be host-executed. Agent threads and application execution are optional Features, not requirements for a one-off run.