Questions about the draft proposal on interrupts

[pitt-liang]

https://docs.ag-ui.com/drafts/interrupts

1. In the current proposal, the run_finished event appears to deliver only a single interruption. How does the design support multiple interruptions within a single run?

For example, interruptions may occur at multiple nodes in a workflow, especially when some nodes are running in parallel. Another scenario is when an LLM supports parallel tool calls, and more than one selected tool call requires user confirmation.

example data of run_finished event with interruption.

{
  "type": "RUN_FINISHED",
  "threadId": "t1",
  "runId": "r1",
  "outcome": "interrupt",
  "interrupt": {
    "id": "int-abc123",
    "reason": "human_approval",
    "payload": {
      "proposal": {
        "tool": "sendEmail",
        "args": { "to": "a@b.com", "subject": "Hi", "body": "…" }
      }
    }
  }
}

2. What is the current status or progress of this draft proposal? I noticed that the last update was about three months ago.

Thanks.

[NathanTarbert]

Hi @pitt-liang, I looked into this and we are planning on implementing interrupts in the new year.

[iismd17]

Proposal: Multi-Interrupt Support

Hey folks, I know you are about to pick up this task soon and I wanted to suggest an update to the proposal.

Building on @pitt-liang's question - the current draft handles single interrupts well, but real-
world agent workflows often need multiple simultaneous approvals (parallel tool calls, batch
operations, multiple workflow nodes pausing at once).

The Problem

When an LLM makes multiple tool calls in one turn, each requiring approval, the current single-
interrupt model forces sequential handling - N tools means N interrupt/resume cycles. This
creates poor UX (sequential dialogs) and unnecessary round-trips.

Proposed Extension

1. Add interrupts array to RunFinished:

type RunFinished = {
  type: "RUN_FINISHED"
  outcome?: "success" | "interrupt"
  result?: any

  // Existing draft single interrupt - could be replaced if we don't care about backwards compatibility with draft proposals
  interrupt?: {
    id?: string
    reason?: string
    payload?: any
  }

  // New: multiple interrupts
  interrupts?: Array<{
    id: string           // Required for multi-interrupt
    reason?: string
    payload?: any
  }>
}

2. Add resumes array to RunAgentInput:

type RunAgentInput = {
  // ... existing fields

  // Existing single resume - - could be replaced if we don't care about backwards compatibility with draft proposals
  resume?: {
    interruptId?: string
    payload?: any
  }

  // New: multiple resumes
  resumes?: Array<{
    interruptId: string
    payload?: any
  }>
}

Benefits

• Single round-trip for batch approval of N tools.
• Better UX - users see all pending approvals at once, can approve/deny individually or in
  bulk
• Atomic processing - all decisions collected before execution

This is a minimal extension that solves the parallel tool approval case. We can choose to replace the existing interrupt/resume objects with the respective arrays or keep them around for backwards compatibility - depending on our practices of backwards compatibility requirements with published draft proposals (that I don't know what they are).

[NathanTarbert]

Hi @iismd17, thanks for adding this suggestion.

@mme, would love to hear how you are thinking about this.

[iismd17]

Hey @pitt-liang and @mme sure, I am @kakkarot17 on discord and present in the AG-UI discord. Feel free to reachout to me directly and we can continue the conversation. :)

As I was implementing the draft proposal, I was running into multi-tool call approval situations that draft proposal had not yet added support for which lead me to think that we probably need arrays here instead of single only interrupt to support that use-case - since they solved for the multi-tool call interrupt situation cleanly (unless I am missing some other corner case here too).

[tom13878]

Hi @NathanTarbert,

Do you know if this is being worked on? I couldn't see it here: Road Map

I was planning to implement the draft proposal (for the Python SDK) https://docs.ag-ui.com/drafts/interrupts.

Except that both the RUN_FINISHED event and the RunAgentInput.resume would support multiple interrupts.