[BUG] Using structured_output inside of a Tool responds with: No valid tool use or tool use input was found in the Bedrock response

[braddengross]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

0.1.9

Python Version

3.11.11

Operating System

macOS 15.3.2

Installation Method

pip

Steps to Reproduce

1. Create an Tool with an agent that uses a structured_response of a Pydantic model.
2. Create another agent with access to that tool.
3. The second tool will to use the structured_response tool, but respond with the error: "No valid tool use or tool use input was found in the Bedrock response."

Expected Behavior

The tool should return its normal, expected response. In this case, I took the structured_response and converted into to JSON and returned a normal status/content response.

Actual Behavior

Structured response agent inside of Tool responds with No valid tool use or tool use input was found in the Bedrock response.

Additional Context

This only happens when calling this as a Tool. If I call the @tool decorated directly, it responds correctly.

Possible Solution

It may be that the structured response needs to combine some information from the context that this is a tool invocation.

Related Issues

No response

[zastrowm]

Would you happen to have the code sample that reproduces this?

[theagenticguy]

+1 I'm struggling to understand the complete setup from the description

[theagenticguy]

Possibly related: #307

[braddengross]

This may be an example of user error/expected behavior, but I'd like to get some feedback either way. Here is the simplest example I could make. I think this use case shows the error occurs when the model does not feel it can answer given the prompt. There may be other ways to create it, maybe if the model were to ask a question when expecting a structured response.

You can comment in/out these lines to see the behavior. When given the formula the response should come back as expected, if not the model will ask a follow up question and break. I understand that the expected behavior would likely be to ask for a structured response when you're confident you have the information to create one. I believe I may be seeing this case in other areas, but the replication is too complex.

Lines to comment in and out to run with/without the issue:

    # prompt=request.prompt + " " + formula.formula,
      prompt=request.prompt,

Code:

import boto3
from fastapi import FastAPI
from pydantic import BaseModel
from strands import Agent, tool
from strands.models import BedrockModel

app = FastAPI()

boto_session = boto3.session.Session(region_name="us-east-1")
claude_model_id = "us.anthropic.claude-3-7-sonnet-20250219-v1:0"

bedrock_model = BedrockModel(
    model_id=claude_model_id,
    temperature=0.0,
    boto_session=boto_session,
)


@tool
def formula_namer_tool(formula: str) -> dict:
    boto_session = boto3.session.Session(region_name="us-east-1")

    bedrock_model = BedrockModel(
        model_id=claude_model_id,
        temperature=0.0,
        boto_session=boto_session,
    )

    feature_namer_agent = Agent(
        system_prompt=f"""
            You are a helpful assistant that helps users understand features in their data.
            Given a feature formula, you will create a name and description for the feature.
            The response MUST be valid JSON with the format:
            "{{feature_actions" : [{{"name": <SUGGESTED NAME>, "description": <SUGGESTED DESCRIPTION>}}]}}
            """,
        model=bedrock_model,
    )
    query = f"""
            Feature formula: {formula}
            """
    response = feature_namer_agent.structured_output(
        output_model=FormulaRecommendation, prompt=query
    )
    json_response = response.model_dump_json()

    return {
        "status": "success",
        "content": [
            {
                "text": f"✅ Created this feature name and description: {json_response}",
            }
        ],
    }


agent = Agent(
    system_prompt="Given a formula, use tools available to you to create a name and description",
    model=bedrock_model,
    tools=[formula_namer_tool],
)


class FormulaRecommendation(BaseModel):
    name: str
    desc: str


class Formula(BaseModel):
    formula: str


class FormulaNamerRequest(BaseModel):
    formulas: list[Formula]
    prompt: str


class FormulaNamerResponse(BaseModel):
    formula_recommendations: list[FormulaRecommendation]


@app.post("/test_endpoint")
async def name_feature(request: FormulaNamerRequest):
    responses = []
    for formula in request.formulas:
        response = agent.structured_output(
            output_model=FormulaRecommendation,
            # prompt=request.prompt + " " + formula.formula,
            prompt=request.prompt,
        )
        responses.append(response.model_dump_json())
    return responses

Request:

{
    "prompt": "Create names and descriptions for features that are missing name and/or description",
    "formulas": [
        {
            "formula": "if(is_home, 1.1, 1.0)"
        },
        {
            "formula": "if(is_away, 1.0, 1.1)"
        }
    ]
}