Introduce OpenAIAgent backed by the Response API in Extensions.

[ekzhu]

Confirmation

• I confirm that I am a maintainer and so can use this template. If I am not, I understand this issue will be closed and I will be asked to use a different template.

Issue body

Assistant API will be deprecated in 2026, replaced by Response API: https://platform.openai.com/docs/guides/responses-vs-chat-completions.

We should start support the Response API by introducing OpenAIAgent that is backed by the Response API, which can be stateful.

The new agent should conform to the behavior of AssistantAgent in AgentChat, while backed directly by the openai library.

Thoughts, feedback welcome!

[federicovilla55]

Should OpenAIAgent contain an agent implementation that uses the Response API to generate responses or should it provide a common abstraction layer that unifies both the Assistant API and the Responses API under a single interface? (and define in a new class the implementation of the Response API)
Since the Assistant API will be deprecated in 2026, only the Responses API will remain, but introducing a common interface could simplify the transition (as with a common interface developers could select the appropriate adapter based on their configuration or runtime and therefore it's easier to switch).

Regarding the agent implementation that uses Responses API to generate model responses:

The context in Response API can be maintained using a previous_response_id that is updated on every response (as this API has stateful interactions). This ensures that instructions from previous responses are not unintentionally carried over to subsequent ones.
Since the state model should be saved and reloaded between session (e.g. for resuming a session after a restart), the state model should be serializable, for example by using Pydantic’s BaseModel. I think the same should hold for the file IDs from the files uploaded via the OpenAI’s API. A class equivalent to OpenAIResponseAgentState(BaseModel) could be implemented for such scope.

Additionally, most of the logic from the current Assistant API implementation can be adapted for the Responses API, with necessary modifications to accommodate its different response format; therefore much of this functionality could be leveraged from the existing OpenAIAssistantAgent class within the autogen_ext.agents.openai module.