Agent that has Agent `asTool` that has Tool that `needsApproval` does not return interruptions

[Juicetan]

Please read this first

• Have you read the docs? Agents SDK docs ✅
• Have you searched for related issues? Others may have faced similar issues. ✅

Describe the bug

When I have an Agent as a tool that has a tool itself that has needsApproval to be true, the run result.interruptions array remains empty.

Debug information

• Agents SDK version: v0.0.11
• Runtime environment Node.js 22.15.1

Repro steps

Agent A has Agent B asTool in tools array

const agentA = new Agent({
  ...
  tools: [
    agentB.asTool({
      ...
    })
  ]
})

Agent B has Tool A in tools array

const agentB = new Agent({
  ...
  tools: [
    toolA
  ]
})

Tool A has needsApproval to return true

const toolA = tool({
  ...
  needsApproval: async (_ctx, _input) => {
    console.log('> approval check', _ctx, _input);
    return true;
  }
})

Run result does not return any interruptions

const result = await runner.run(agentA, conversation, promptOpts);
if(result.interruptions?.length){
  console.log('> pause for approval', JSON.stringify(result.state, null, 2));
}

Expected behavior

Expected behavior is that result.interruptions will have items, however it does not. The conversation just returns a finalOutput that implies that something is happening in the background.

What I have noticed is that if I have Agent B has a handoff within Agent A, then it works and I receive result.interruptions items, however its conversational behavior as a handoff is not what I'm trying to achieve.

Is this intended SDK behavior?

[seratch]

Thanks for taking the time to report this. I will look into it later this week.

Just to confirm, do you mean that even when toolA is called, HITL does not work at all and interruptions is empty in your code?

If toolA is not called, that could be an issue with your instructions and/or the clarity of your tool description.

[Juicetan]

yes i can confirm that toolA is invoked as i have log statements within the execute function and the run result has an empty interruptions array.

[vrtnis]

One way to address this, I suppose, is to add a returnRunResult option that lets callers access the full RunResult, including any interruptions like approval requests. Right now, asTool() always returns just the final message string. So if a sub-agent calls a tool that requires approval (needsApproval: true), the outer agent never sees the interruption.

The returnRunResult flag (see #249) is totally opt-in and doesn’t change existing behavior. But for devs building nested agent workflows, esp those involving approvals, it enables proper pause and resume behavior without losing control or having to restructure the flow. Thoughts?

Also, this is perhaps conceptually similar to the stream_inner_events enhancement proposed in the python sdk (openai/openai-agents-python#864), which allows inner agent tool calls and events to be surfaced.

Here we offer a simpler way via returnRunResult instead of streaming but the motivation is closely aligned.

Would love feedback on the draft PR. Happy to iterate!

[github-actions]

This issue is stale because it has been open for 7 days with no activity.

[github-actions]

This issue was closed because it has been inactive for 3 days since being marked as stale.

[Juicetan]

@seratch is this going to be discarded as a non-issue? i believe this coincides with core behavior