Audience: developers choosing a local, CI/CD, or long-lived runtime mode.
Create an AgentRuntime with the target Conductor configuration. Ensure every
Python @tool is importable in the process that will call serve() or run().
| Operation | Effect |
|---|---|
runtime.plan(agent) |
Compile only; does not register or execute. |
runtime.deploy(agent) |
Compile and register; does not execute. |
runtime.serve(agent) |
Deploy, start local tool workers, and block. |
runtime.run(agent, prompt) |
Start and wait for an AgentResult. |
runtime.start(agent, prompt) |
Start and return an AgentHandle. |
runtime.resume(execution_id, agent) |
Reattach workers to a prior execution. |
Use deploy in CI/CD and serve in long-lived worker processes. run is the
best local quickstart. Always call shutdown() or use with AgentRuntime() for
short-lived programs. See the runtime reference.
Compile/register with deploy() during release, then run serve() in one or more
long-lived worker services. Use plan() in CI to inspect the compiled workflow
before deployment. Use start() for asynchronous callers and resume() after a
local worker process restart.
deploy() returns deployment information without executing an agent; serve()
blocks until interrupted; run() returns a terminal AgentResult. Use the
runtime context manager or shutdown() to stop local polling cleanly.
See deployment and scaling, runtime reference, and agent client.