Resolving the Duplicate 'prompt_tokens' Merge Error in LangGraph PreBuilt ReAct Agents

Below is a possible explanation and some workarounds based on the LangChain documentation and known issues:

---

What’s Happening

The error message

  TypeError: Additional kwargs key prompt_tokens already exists in left dict and value has unsupported type <class 'int'>

indicates that during the merge of metadata dictionaries (for example, merging run metadata or additional kwargs coming from calls to multiple ReAct agents), the merge utility in LangChain is trying to combine values for the key "prompt_tokens." In your case, the value is an integer—which the merge code isn’t set up to handle (it expects types like dictionaries or lists that can be combined).

This is similar to the issue reported for Gemini (see #23827). Although Agent 1 works fine, Agent 2 ends up triggering this merge behavior because both agents output metadata with the same key, but with numeric values that cannot simply be “added” together.

---

Possible Workarounds

1. Upgrade or Patch the Library:
   Check if you are using the latest version of LangChain and its dependencies (langchain_core, langgraph, etc.). A recent update might have a fix for this merging behavior. If not, you might consider temporarily patching the merge functionality in _merge.py to either override or skip merging numeric values for keys like prompt_tokens.

2. Isolate Agent Metadata:
   If possible, configure your agents so that the metadata output (especially additional kwargs) does not include keys that conflict when merging. For example, if you can remove or rename the “prompt_tokens” key from one of the agents’ output dictionaries, the merge utility will no longer encounter duplicate keys with unsupported types.

3. Customize the Agent’s Output Parser or State Modifier:
   Since you’re using prebuilt ReAct agents with different prompts, consider adding a state modifier or output parser that cleans or normalizes the metadata before the merging step is encountered by LangGraph. By ensuring that numeric token counts do not get merged (or are transformed into a different structure, such as wrapping them in a dict), you may be able to work around the merge issue.

---

Example Patch (Conceptual)

If you wish to experiment with a quick patch in the merge function (note that this is only for testing purposes and should be done with caution), you might modify the inner logic of merge_dicts so that when encountering an int you simply override rather than attempt to add:

PYTHONCopied
def merge_dicts(left, *others):
    merged = left.copy()
    for right in others:
        for right_k, right_v in right.items():
            if right_k not in merged:
                merged[right_k] = right_v
            else:
                # If the key exists with an int, override it
                if isinstance(merged[right_k], int) and isinstance(right_v, int):
                    merged[right_k] = right_v
                elif isinstance(merged[right_k], dict):
                    merged[right_k] = merge_dicts(merged[right_k], right_v)
                elif isinstance(merged[right_k], list):
                    merged[right_k] = merged[right_k] + right_v
                else:
                    msg = (f"Additional kwargs key {right_k} already exists in left dict and "
                           f"value has unsupported type {type(merged[right_k])}.")
                    raise TypeError(msg)
    return merged

This patch simply overrides the value for keys that are integers instead of trying to merge them.

---

Conclusion

The merge error arises from a conflict in additional kwargs—specifically the prompt_tokens key—when using two identical ReAct agents with different prompts inside a LangGraph graph. You can work around this by:

• Upgrading your libraries if a fix exists.
• Adjusting the agents to prevent duplicate numeric metadata keys.
• Applying a custom patch in the merge utility to handle this scenario.

I hope this helps you troubleshoot and resolve the issue!