Agent responses are repeated multiple times

[raul-marcano-reimundez]

Describe the bug

Sometimes an agent response is repeated multiple times within the same turn.

This leads to two main problems:

• Performance impact: the request takes significantly longer compared to other calls.
• Cost impact: the repeated responses are then included as input in the following conversation turn, unnecessarily increasing token usage and costs.

Debug information

• Agents SDK version: v0.3.2
• Python version Python 3.11.9

Repro steps

Run an agent conversation using the Python SDK.
Observe that in SOME CASES the assistant’s reply is duplicated or repeated multiple times.
Check the following conversation turn and note that the duplicated messages are included as part of the input.

Expected behavior

The agent should return its message once per turn, without duplication.

Screenshots / Logs:

Attached are screenshots from the OpenAI platform showing the issue.

[seratch]

Thanks for writing in. If you could share code snippet demonstrating how your agent works and/or the actual traces showing how the agent loop went, we may be able to better understand how things went. Basically, this kind of issue could be mitigated by making your instructions and tool calling even clearer for LLMs.

[viniciusarruda]

Hi, I'm also encountering this issue. Sometimes the duplicate assistant messages are identical, and other times they differ.

I understand that prompt design can help mitigate this, but in my case, the problem occurs even when using structured outputs. From the documentation, my understanding is that the loop should stop once the output matches the schema of the specified output type. In my situation, the schema is matched, yet duplicate assistant messages are still generated -- suggesting that the loop is not being stopped as expected.

This makes me wonder: is the issue with the Agents SDK not detecting the output type properly, or is the LLM itself failing to signal completion (e.g., not producing the end-of-token when it should)?

I'm wondering if it makes sense to check new_items and filter out duplicate assistant messages at the end of the list.

[viniciusarruda]

This issue has been reported before: https://community.openai.com/t/prevent-multiple-assistant-messages-in-a-run/932129.