State not updating

[yamirghofran]

When running a reduced version of our tool-caller agent workflow I noticed what seems like an error in the state handling between tool calls. The run behavior is like this:

1. the router works (with manual routing logic), the ignored LLM output looks fine,
2. the first tool call run_mpnn appears to work
3. The following tool step -- score_mpnn -- fails to update the state. As a result, score_mpnn runs again for some or all pipelines, maybe looping on that tool a time or two and crashing.
   That all happens in about two minutes on a single A100 after which the job hangs.

I can't tell exactly where is the disconnect between the successful tool run and the failed state update, but thought maybe you could take a look? My code, Anvil slurm script, and example slurm output are attached.

mason_yesterdayupdate.py

[yamirghofran]

I believe I have found the issue.

The tools were wrapped with AsyncFlowType.AGENT_TOOL_AS_FUNCTION, which creates LangChain tools. LangChain tools return values that get added to messages/state as results, not state updates. When the tool implementation returned a complete state dictionary like:

return {
    **state,
    "tools_call_count": tools_call_count + 1,
}

This return value was treated as a tool result, not a state update, so LangGraph didn't merge it into the state.

I've attached the fix to your code which changes the tool wrappers from AGENT_TOOL_AS_FUNCTION to EXECUTION_BLOCK.

In this case, the EXECUTION_BLOCK wrapper:

1. takes the full state as input
2. returns state updates that LangGraph automatically merges into the current state
3. is designed for nodes that modify state (like the router, which already used this type correctly)

This ensures that when tools return updated state dictionaries, LangGraph properly merges them into the workflow state.

mason_yesterdayupdate_fixed.py