agent.invoke() with Gemini gives inconsistent output types

Question

I am using the LangChain 1.0 new agent = create_agent() method to create an agent but I am getting inconsistent output response types.

Case 1: str as output When the user query is simple such as "What is a banana" the content property returns a string type:

response3 = agent.invoke(
    {"messages": [{"role": "user", "content": "What is a banana?"}]},
    {"configurable": {"thread_id": "2"}} 
)

print(response3["messages"][-1].content)

Output:

A banana is an elongated, edible fruit botanically a berry, produced by several kinds of large herbaceous flowering plants in the genus Musa.

Case 2: list as output But if the user query is confusing such as "What's its name" the content property returns a list type:

response3 = agent.invoke(
    {"messages": [{"role": "user", "content": "What's its name?"}]},
    {"configurable": {"thread_id": "2"}}  
)

print(response3["messages"][-1].content)

The output is:

[{'type': 'text', 'text': 'I\'m sorry, I don\'t understand what "it" refers to. Could you please provide more context?', 'extras': {'signature': 'CscCAdHtim+SJIpPCDrUbhw9W'}}]

This happens only when I am using gemini-2.5.-flash. It does not happen with openai models.

This inconsistency would cause unexpected bugs. Is there a proper way to handle this without doing type checks with conditionals?

Edit: This seems to happen only with gemini-2.5-flash. I tried gemini-2.0-flash and I got consistent str types.

Ajeet Verma · Accepted Answer · 2025-11-18 01:08:24Z

The issue you're experiencing occurs because Gemini's response format sometimes includes structured metadata alongside text content, and LangChain preserves that native structure.

When the model adds safety-related metadata (like the signature field), it returns the full list rather than just extracting the text string.

The good news is that LangChain v1 provides a standardized way to access content that abstracts away provider differences. To make the output uniform across the models used, you can try AIMessage.content_blocks instead.

response = agent.invoke(
    {"messages": [{"role": "user", "content": "What's its name?"}]},
    {"configurable": {"thread_id": "2"}}
)

# This handles Gemini, OpenAI, and other models consistently
for block in response["messages"][-1].content_blocks:
    if block.type == "text":
        print(block.text)  # Always works

The content_blocks property standardizes the output across all providers and normalizes Gemini's native format to a consistent interface.

Collectives™ on Stack Overflow

agent.invoke() with Gemini gives inconsistent output types

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related