Audience: applications owning local tool workers and Conductor-agent lifecycle.
Create one runtime per application lifetime with a configured server connection. Use a context manager for scripts and an application shutdown hook for services.
AgentRuntime owns connection-facing clients, local tool workers, and execution
lifecycle. Use it as a context manager or call shutdown().
| Method | Returns | Purpose |
|---|---|---|
run / run_async |
AgentResult |
Start and wait for completion. |
start / start_async |
AgentHandle |
Start without waiting. |
stream / stream_async |
stream | Consume agent events. |
plan |
compiled definition | Compile without registration or execution. |
deploy / deploy_async |
deployment information | Register agent definitions. |
serve |
none | Deploy and poll tool workers. |
resume / resume_async |
AgentHandle |
Reattach to an execution. |
AgentConfig.from_env() reads canonical CONDUCTOR_AGENT_* settings. Connection,
auth, and SDK log level belong to Configuration and use CONDUCTOR_*. Share one
runtime for an application lifetime; do not create one per request.
RunSettings overrides model/runtime options for one run without mutating the
agent definition. Use on_event or stream() for progress, and resume() to
reattach local workers after a process restart. A SCHEDULED tool task means no
compatible worker process is polling; a model error normally means the server
provider setup is incomplete.
Call shutdown() or exit the context manager to stop local workers. Continue with
runtime modes and client control.