Skip to content

[Layer 6] Should negotiation support a mid-task pause — as MCP and A2A both now do? #6

Description

@marc-poljak

Context. Layer 1 — A2A Binding § Open Questions: "A2A's resumable-interrupt states could plausibly give Layer 6 (negotiation) a native 'task paused, awaiting counter-offer' mechanism. This is a Layer 6 question, not a Layer 1 one — noted here as a pointer, not resolved."

Today, Layer 6 negotiation happens entirely before acmp/invoke: offer, counter-offer, accept, then execution starts. Once a task is running there is no way to pause and renegotiate — a provider that discovers mid-task that the job is bigger than quoted can only finish at a loss, fail, or be cancelled.

What changed. Both of ACMP's candidate transport substrates now model exactly this pause:

  • A2A has the input-required / auth-required task states — non-terminal and resumable.
  • MCP adds Multi Round-Trip Requests in its 2026-07-28 revision: a server returns resultType: "input_required", and the client supplies what's missing by retrying the original request.

Two independent protocols converging on the same shape is a strong hint. It does not follow that ACMP should inherit it.

Your task. Work out whether Layer 6 should gain a paused-renegotiation mechanism, and what it would cost:

  • What may trigger a pause — scope creep, a revised cost estimate, an ambiguity in the input? Or nothing, because that is what max_price_cu and cancellation are already for?
  • What travels in the interrupt: a fresh offer, a delta on the accepted one, or only a reason code?
  • The hard part is economic, not syntactic. A paused task holds escrow while it waits (Layer 4). Who bears the cost of the wait, and what expiry turns a stalled pause into a partial settlement?
  • What stops a provider from using mid-task renegotiation as a hostage tactic, having already consumed the buyer's deadline?
  • Does this need Layer 1 wire support (acmp/interrupt / acmp/resume, see Layer 1 § Open Questions), or can it ride on existing Layer 6 messages?

Done = a comment or PR taking a position on whether ACMP should adopt this at all, with the escrow-economics question at least sketched. A reasoned "no — here is why pre-invoke-only negotiation is the right constraint" is as valuable as a design.

Metadata

Metadata

Assignees

No one assigned

    Labels

    good first issueGood for newcomersquestionFurther information is requestedwg-negotiationLayer 6 — Negotiation Protocol working group

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions