AgentRunner cannot use previous_response_id for multi-turn tool-calls in the first run #2124
Description
Hi there! Hope you’re doing well. This is the version from my pip freeze:
openai==2.8.1
openai-agents==0.6.1
When using openai-agents-python, AgentRunner only leverages previous_response_id if a non-null previous_response_id is passed into the run. This makes it impossible to benefit from Responses API’s incremental conversation state (previous_response_id) inside the very first AgentRunner.run call, even if that run internally performs multiple model turns and tool calls.
In other words, complex “first turns” (multi-tool, multi-round workflows within a single run) cannot use previous_response_id at all, unless the caller already has an existing response id from a previous Responses/Agents call.
Code references
_ServerConversationTracker
@dataclass
class _ServerConversationTracker:
"""Tracks server-side conversation state for either conversation_id or
previous_response_id modes."""
conversation_id: str | None = None
previous_response_id: str | None = None
sent_items: set[int] = field(default_factory=set)
server_items: set[int] = field(default_factory=set)
def track_server_items(self, model_response: ModelResponse) -> None:
for output_item in model_response.output:
self.server_items.add(id(output_item))
# Update previous_response_id only when using previous_response_id
if (
self.conversation_id is None
and self.previous_response_id is not None
and model_response.response_id is not None
):
self.previous_response_id = model_response.response_id
AgentRunner.run:
previous_response_id = kwargs.get("previous_response_id")
conversation_id = kwargs.get("conversation_id")
...
if conversation_id is not None or previous_response_id is not None:
server_conversation_tracker = _ServerConversationTracker(
conversation_id=conversation_id, previous_response_id=previous_response_id
)
else:
server_conversation_tracker = None
So:
• _ServerConversationTracker is only created when conversation_id or previous_response_id is not None.
• And track_server_items only updates previous_response_id when it is already non-None
This means:
• If I call Runner.run(..., previous_response_id=None, conversation_id=None) for my first ever turn:
• server_conversation_tracker = None
• Every internal model call in that run sends the full input items (original prompt + all generated items so far).
• previous_response_id is never used or updated within that run.
Why this feels wrong
previous_response_id mode only works if you already have a prior response id, i.e., it’s only useful to chain separate runs, not to optimize the internal multi-turn workflow of a single first run.
From a user perspective, this might be surprising and limiting:
• A single AgentRunner.run might contain many internal turns (model → tool → model → …).
• For a complex first user question, this first run can be the heaviest one.
• But there is no way to say: “for this first run, please start with no prior state, but after the first LLM call, keep using its response_id as previous_response_id for subsequent internal turns”.
Expected behavior
I would expect at least one of the following to be possible:
- First run can automatically “enter” previous_response_id mode after the first model call.
- Even if the caller does not pass a previous_response_id, the first internal get_response in the run could seed previous_response_id for later turns in the same run.
- Or: The SDK exposes a clear way to opt in to using previous_response_id within a run, even when starting from no previous response.
- For example, the presence of some flag or config could tell AgentRunner to create a _ServerConversationTracker and to start using response_id of the first internal call as previous_response_id for subsequent calls.
Thanks for your help.