What causes "DeepAgentRuntime must be started before invoke" in langchain-ai/deepagents?

Call `await runtime.start()` once after construction and before any invoke

langchain-ai/deepagents · error · RuntimeError

DeepAgentRuntime must be started before invoke

Error message

DeepAgentRuntime must be started before invoke

What it means

DeepAgentRuntime builds its agent graph lazily in start(); invoke() raises this RuntimeError when self._graph is still None, i.e. the runtime was constructed but never started. The graph cannot be invoked before the async startup completes.

Source

Thrown at libs/talon/deepagents_talon/runtime.py:329

            result = cleanup()
            if isinstance(result, Awaitable):
                await result

    async def invoke(self, request: AgentRequest) -> AgentResult:
        """Invoke the Deep Agents graph for one Talon request.

        Args:
            request: Agent request supplied by the Talon host.

        Returns:
            Final assistant text from the graph.

        Raises:
            RuntimeError: If the runtime has not been started.
        """
        if self._graph is None:
            msg = "DeepAgentRuntime must be started before invoke"
            raise RuntimeError(msg)

        token = _CRON_ORIGIN.set(_cron_origin_from_request(request))
        try:
            text = await self._invoke_until_text(request)
        finally:
            _CRON_ORIGIN.reset(token)
        return AgentResult(text=text)

    def _build_tools(self) -> list[BaseTool | Callable[..., object]]:
        tools: list[BaseTool | Callable[..., object]] = []
        if self.include_web_tools:
            tools.extend([fetch_url, web_search])
        if self.cron_store is not None:
            cron = CronTools(store=self.cron_store, origin=_current_cron_origin)
            tools.extend(cron.as_langchain_tools())
        tools.extend(self.tools)
        return tools

View on GitHub (pinned to a1af029e6e)

Solutions

Call `await runtime.start()` once after construction and before any invoke
If start() previously raised, fix the startup error and construct/start a fresh runtime rather than reusing the half-initialized one
Wrap the start-then-invoke sequence in your app's lifecycle management (e.g. FastAPI lifespan) so it always runs

Example fix

// before
runtime = DeepAgentRuntime(...)
reply = await runtime.invoke(request)
// after
runtime = DeepAgentRuntime(...)
await runtime.start()
reply = await runtime.invoke(request)

Defensive patterns

Strategy: try-catch

Try / catch

try:
    reply = await runtime.invoke(request)
except RuntimeError as exc:
    if "must be started before invoke" in str(exc):
        await runtime.start()
        reply = await runtime.invoke(request)
    else:
        raise

Prevention

Always pair construction with `await runtime.start()` in a lifecycle hook (app startup, fixture setup)
Never reuse a runtime after teardown; build a fresh one
Wrap start+invoke in a small helper so callers cannot skip start

When it happens

Trigger: Calling `await runtime.invoke(request)` without first doing `await runtime.start()`, or calling invoke after an explicit teardown/stop that cleared the graph.

Common situations: Forgetting to await start() in a script or test fixture; a failed start() left the graph unset and a later retry calls invoke directly; reusing a runtime object after shutdown.

Related errors

AI-assisted analysis of langchain-ai/deepagents@a1af029e6e (2026-08-29). Data as JSON: /api/errors/b8d6b2ca8b981261. Report an issue: GitHub.