OpenAI Agents SDK

The OpenAI Agents SDK owns the paused agent state. ActionBox provides a durable request for the human decision; your application still owns the tool's credentials and side effect.

Ownership boundary

ActionBox is a decision service, not a replacement for the Agents SDK state store. Persist a resumable state in your own application when the approval wait can outlive the worker that created it.

Supported flow

1. Mark only consequential tools with needs_approval=True.
2. When a run returns result.interruptions, create one ActionBox Action per interruption with a stable tool-call ID and bounded review context.
3. Wait for the typed boolean response. Do not treat a missing response, a string, or a local timeout as approval.
4. Call result.to_state(), then apply each decision to that exact interruption with state.approve(...) or state.reject(...).
5. Resume the original state with Runner.run(agent, state).
6. Run the approved tool, then optionally report its real execution result to ActionBox. A decision authorizes an attempt; it does not prove success.

Install the public SDKs in the environment that runs the agent:

python -m pip install actionbox-sdk openai-agents

The state transition is intentionally applied to the original result:

from typing import Any

from agents import Agent, Runner


async def resume_approved_run(
    agent: Agent,
    result: Any,
    decisions: list[bool],
):
    state = result.to_state()
    for interruption, approved in zip(result.interruptions, decisions, strict=True):
        if approved:
            state.approve(interruption, always_approve=False)
        else:
            state.reject(
                interruption,
                rejection_message="Rejected or timed out in ActionBox.",
            )
    return await Runner.run(agent, state)

Use a stable tool-call ID in the ActionBox idempotency key. If a worker retries the same interruption, it should recover the same request instead of creating a second notification. Handle every interruption in the result, and use always_approve=False so one approval does not silently authorize future calls.

For a complete walkthrough, see OpenAI Agents SDK human-in-the-loop approval.