[Feature Request] OpenAI Agents Streaming support

[surki]

Feature Request

1. Looks like run_streamed() is not supported. In a long running reasoning agent, we need to keep the chat updated with the progress. Any plan to support it?

Answered Questions

2. What's the recommended way to integrate? the OpenAIAgentsPlugin or set_open_ai_agent_temporal_overrides? If plugin, any plan to cut a release with latest changes?
3. Is it safe to use single worker for both regular workflows/activities and agent workflows? concerned about plugin/override negatively interacting with other workflows.

[tconley1428]

1. We do plan to support streaming. It's on the roadmap.
2. OpenAIAgentsPlugin is intended to be the only usage method. It released last night: https://github.com/temporalio/sdk-python/releases/tag/1.15.0, with samples and readme updates in the works to align to that new model.
3. It is safe to use both. You can look at the plugin implementation which is quite small to identify the set of changes. The only one of particular concern is that other workflows will have to use the pydantic_data_converter. This is generally not a bad idea anyway since it is more robust than our built in one, but for some scenarios it could be problematic. For instance, adding it into an existing worker with histories may not work as the conversion process would differ.

I'll repurpose this issue specifically to track the streaming support request. In general for free form questions rather than a feature request, I would recommend using our community slack: https://temporalio.slack.com/archives/CTT84RS0P

[surki]

Thanks @tconley1428 , I will keep an eye on this for streaming support, happy to test when it lands.

I upgraded and moved to use plugin, seems good so far, didn't notice any issues, will test some more.
I didn't see any examples for doing pytest, seems temporalio.testing.WorkflowEnvironment doesn't support plugins?

I seem to be using non-pydantic version for other workflows, so would run two workers for now.

Thanks for sharing community slack, will use that going forward for the questions.

[tconley1428]

For pytest, you can see https://github.com/temporalio/sdk-python/blob/main/tests/contrib/openai_agents/test_openai.py. In our particular case the WorkflowEnvironment was already generated, so adding it there didn't come up, but that's something we can do.

[mav913ai]

I am new to Temporal - just was trying to test it with OpenAI agents sdk, but couldn't since Runner.run_streamed is not supported. Has anyone figured out a workaround for this?

We can't run the agent without streaming.

[tconley1428]

Streaming is not currently supported, as it is non-trivial to bring those updates back to the workflow from the activity in which the models are running. It is on the roadmap, but not currently supported.

[brianstrauch]

Streaming is now supported for the OpenAI Agents SDK integration, but experimental. See here: https://docs.temporal.io/develop/python/integrations/openai-agents#streaming