400 error when using saved prompt with built-in tool names in tool_choice

[fathulfahmy]

Describe the bug

When using Agents SDK with reusable prompt, tools defined in prompt is not found. Workaround is to define tool in source code.

Debug information

• Agents SDK version: v0.2.3
• Python version: Python 3.11

Repro steps

1. Create a prompt in OpenAI dashboard.

Model
name: gpt-4.1-mini
text.format: text
tool_choice: web_search_preview
temp: 0.00

Tools
Web Search

System message
Search reputable web sources to verify and gather information about the customer's occupation and its typical stability, income potential, or risk level.

**Example:**

Reasoning:  
The customer is a software engineer, a profession typically associated with high job stability and above-average income [Bureau of Labor Statistics]. The technology industry has displayed consistent growth and demand for experienced developers [Indeed Hiring Lab]. 

References:  
- https://www.bls.gov/oes/current/oes151252.htm  
- https://www.hiringlab.org/2023/01/10/tech-job-trends/  

2. Copy prompt id e.g. pmpt_1234567890.
3. Attach prompt id to an agent.

    OPENAI_JOB_SEARCH_PROMPT_ID = "pmpt_123"

    job_search_agent = Agent(
        model="gpt-4.1-mini",
        name="job_searcher",
        prompt={"id": OPENAI_JOB_SEARCH_PROMPT_ID},
    )

    job_search_result = await Runner.run(job_search_agent, "Software Developer in Malaysia")

    return job_search_result.final_output

Error:

raise self._make_status_error_from_response(err.response) from None
openai.BadRequestError: Error code: 400 - {'error': {'message': "Tool choice 'web_search_preview' not found in 'tools' parameter.", 'type': 'invalid_request_error', 'param': 'tool_choice', 'code': None}}

Workaround:

    OPENAI_JOB_SEARCH_PROMPT_ID = "pmpt_123"

    job_search_agent = Agent(
        model="gpt-4.1-mini",
        name="job_searcher",
        prompt={"id": OPENAI_JOB_SEARCH_PROMPT_ID},
        tools=[
            WebSearchTool(
                search_context_size="high",
                user_location={
                    "type": "approximate",
                },
            )
        ],
    )

    job_search_result = await Runner.run(job_search_agent, "Software Developer in Malaysia")

    return job_search_result.final_output

Expected behavior

Tools to be managed and called using reusable prompt.

    OPENAI_JOB_SEARCH_PROMPT_ID = "pmpt_123"

    job_search_agent = Agent(
        model="gpt-4.1-mini",
        name="job_searcher",
        prompt={"id": OPENAI_JOB_SEARCH_PROMPT_ID},
    )

    job_search_result = await Runner.run(job_search_agent, "Software Developer in Malaysia")

    return job_search_result.final_output

[seratch]

Thanks for reporting this issue. You're right that this is a misbehavior of this SDK.

When you use Responses API directly, no need to pass tools:

from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    prompt={"id": "pmpt_..."},
    input="What's the largest animal in the world?",
)
print(response.output_text)

However, when using the same prompt ID in agent apps, it fails to send a request towards Responses API:

from agents import Agent, Runner, WebSearchTool, Prompt

agent = Agent(
    name="My Agent",
    model="gpt-4.1-mini",
    prompt=Prompt(id="pmpt_..."),
    # Having tools is necessary to make this functional
    tools=[WebSearchTool()],
)
response = Runner.run_sync(agent, "What's the largest animal in the world?")
print(response.final_output)

[fathulfahmy]

@seratch
Thank you for the response. The ability to orchestrate multiple agents and run in parallel is what we're looking for and Agents SDK is the perfect fit for it. Only missing feature for us is to manage the prompts without touching the source code.

Great work Agents SDK team and contributors! Looking forward to the enhancement.