Skip to content

Design exploration: one task entry point with agent claiming and directly routed user questions #355

Description

@Tryanks

Idea

Explore a single user-facing entry point where identifiable model/harness participants can volunteer for a task individually or as a team. Their questions and the user's answers appear in that same surface. This is a product/design exploration; the final interaction may be a task workspace rather than a conventional chat.

Example: a user submits a broad research-heavy request with little initial detail. Fable claims it and consults Astra. If both decide to question the user, each participant's questions appear directly in chronological order in the main surface. Answering an Astra question routes the answer to Astra's originating request; Astra does not first tell Fable to ask on its behalf.

Current foundation

Orchestrate already separates execution dispatch from peer collaborate, provides child threads and callbacks, and supports structured user-input requests. Current reporting is child-to-initiator, with the report explicitly serving as the initiator's view of the child's work. Task claiming and a shared stream of independently routed participant questions are additional semantics, not a rename of dispatch or callback rendering.

Design questions and deliverables

  • Define participant identity (model, harness, profile, effort and personality), eligibility, claiming/arbitration, team formation, and who remains accountable for delivery.
  • Design attributed questions and answers, stable request routing, chronological ordering, multiple outstanding questions, cancellation and late answers. Decide whether answers to related questions are shared and make that visible.
  • Preserve user control when agents disagree, duplicate a claim, need clarification, or a participant disappears.
  • Produce an interaction storyboard or prototype of the Fable/Astra example and a minimal event/state contract, including provider limitations. Prefer existing structured request IDs and thread ownership where they fit.

Exit criteria

A human reviews the proposed interaction. The prototype demonstrates two agents asking directly, interleaved answers delivered to the correct origin, and recovery from an abandoned request without requiring lead-model relaying. No commitment to autonomous claim policy or a final UI layout is implied by this issue.

Code context

Reviewed against main at b4c549787da6 (2026-09-08).

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions