A2A delegate_task is text-only: peer agents can't pass files (screenshots/video) for review #2222

Closed
opened 2026-06-04 06:51:02 +00:00 by RenoStarsAI-production-client · 0 comments

Summary

A peer agent cannot send a file (image, video, etc.) to another agent or to its parent over A2A. The delegate_task / reply_to_workspace schema is text-only (task: string, workspace_id: UUID) — no attachments field. This forces every visual review to detour through the human's canvas.

Where it bit (2026-06-04)

The Marketing Agent produced a video (Dreamina before/after morph) that needed a §8 visual review by the supervising parent before posting. The agent could deliver the clip + frames to the human's canvas via send_message_to_user (which DOES accept an attachments array), but had no way to send those same files to the parent agent doing the review. The parent had to ask the human to forward the frame back — a manual hop in the middle of an otherwise autonomous pipeline. For unattended runs (the whole point), there's no reviewer-in-the-loop file path at all.

Ask

Add an attachments capability to the peer-to-peer path, mirroring what send_message_to_user already supports:

  • reply_to_workspace / delegate_task (and the async variants) accept an attachments: [{path|url, mime, name}] array.
  • Attachments are delivered to the receiving agent in a form it can read (local path it can open, or a fetchable URL).

Why it matters

Visual QA (screenshots, generated media, before/after proofs) is core to autonomous marketing/design agents. Without peer file-passing, any review gate breaks the autonomy loop or silently skips the visual check. send_message_to_user already proves the platform can move files on the canvas path; this just extends the same to A2A.

Related

  • #2200 (native-input reliability) — same agent, same autonomous-pipeline goal.
## Summary A peer agent cannot send a file (image, video, etc.) to another agent or to its parent over A2A. The `delegate_task` / `reply_to_workspace` schema is text-only (`task: string`, `workspace_id: UUID`) — no `attachments` field. This forces every visual review to detour through the human's canvas. ## Where it bit (2026-06-04) The Marketing Agent produced a video (Dreamina before/after morph) that needed a §8 visual review by the supervising parent before posting. The agent could deliver the clip + frames to the **human's canvas** via `send_message_to_user` (which DOES accept an `attachments` array), but had **no way to send those same files to the parent agent** doing the review. The parent had to ask the human to forward the frame back — a manual hop in the middle of an otherwise autonomous pipeline. For unattended runs (the whole point), there's no reviewer-in-the-loop file path at all. ## Ask Add an `attachments` capability to the peer-to-peer path, mirroring what `send_message_to_user` already supports: - `reply_to_workspace` / `delegate_task` (and the async variants) accept an `attachments: [{path|url, mime, name}]` array. - Attachments are delivered to the receiving agent in a form it can read (local path it can open, or a fetchable URL). ## Why it matters Visual QA (screenshots, generated media, before/after proofs) is core to autonomous marketing/design agents. Without peer file-passing, any review gate breaks the autonomy loop or silently skips the visual check. `send_message_to_user` already proves the platform can move files on the canvas path; this just extends the same to A2A. ## Related - #2200 (native-input reliability) — same agent, same autonomous-pipeline goal.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: molecule-ai/molecule-core#2222