What causes "Runtime is already running." in microsoft/autogen?

Track started state yourself and only call start() once per lifecycle

microsoft/autogen · error · ValueError

Runtime is already running.

Error message

Runtime is already running.

What it means

GrpcWorkerAgentRuntime.start() is not idempotent: it checks the internal _running flag and raises ValueError('Runtime is already running.') if start() is called again without an intervening stop(). This protects the single read-loop task and host connection from being duplicated.

Source

Thrown at python/packages/autogen-ext/src/autogen_ext/runtimes/grpc/_worker_runtime.py:264

        self._pending_requests: Dict[str, Future[Any]] = {}
        self._pending_requests_lock = asyncio.Lock()
        self._next_request_id = 0
        self._host_connection: HostConnection | None = None
        self._background_tasks: Set[Task[Any]] = set()
        self._subscription_manager = SubscriptionManager()
        self._serialization_registry = SerializationRegistry()
        self._extra_grpc_config = extra_grpc_config or []
        self._agent_instance_types: Dict[str, Type[Agent]] = {}

        if payload_serialization_format not in {JSON_DATA_CONTENT_TYPE, PROTOBUF_DATA_CONTENT_TYPE}:
            raise ValueError(f"Unsupported payload serialization format: {payload_serialization_format}")

        self._payload_serialization_format = payload_serialization_format

    async def start(self) -> None:
        """Start the runtime in a background task."""
        if self._running:
            raise ValueError("Runtime is already running.")
        logger.info(f"Connecting to host: {self._host_address}")
        self._host_connection = await HostConnection.from_host_address(
            self._host_address, extra_grpc_config=self._extra_grpc_config
        )
        logger.info("Connection established")
        if self._read_task is None:
            self._read_task = asyncio.create_task(self._run_read_loop())
        self._running = True

    def _raise_on_exception(self, task: Task[Any]) -> None:
        exception = task.exception()
        if exception is not None:
            raise exception

    async def _run_read_loop(self) -> None:
        logger.info("Starting read loop")
        assert self._host_connection is not None
        # TODO: catch exceptions and reconnect

View on GitHub (pinned to 027ecf0a37)

Solutions

Track started state yourself and only call start() once per lifecycle
Catch/avoid: check runtime state or guard with try/except ValueError if double-start is possible
Call `await runtime.stop()` (and optionally `await runtime.close()`) before restarting in restart loops

Example fix

# before
await runtime.start()  # ValueError on second call

# after
started = False
if not started:
    await runtime.start()
    started = True

Defensive patterns

Strategy: validation

Validate before calling

# start only if not running (track locally; _running is private)
started = False
async def ensure_started(runtime):
    global started
    if not started:
        await runtime.start()
        started = True

Try / catch

try:
    await runtime.start()
except ValueError as e:
    if "already running" in str(e):
        pass  # idempotent start
    else:
        raise

Prevention

Wrap runtime lifecycle in an async context manager so start() is called exactly once
Avoid retry loops around start() without state tracking
In notebooks, reuse one runtime across cells instead of re-running start()

When it happens

Trigger: Calling `await runtime.start()` twice, e.g. in a retry wrapper, a Jupyter notebook re-execution, or after a framework (like an agent runtime host) already started the runtime for you.

Common situations: Notebook workflows where a cell is re-run; orchestration code that starts the runtime per-task inside a long-lived process; wrapping start() in a generic retry loop.

Related errors

AI-assisted analysis of microsoft/autogen@027ecf0a37 (2026-08-15). Data as JSON: /api/errors/ee832d57b87f5fa1. Report an issue: GitHub.