Have several models from different providers deliberate to a weighted consensus before the agent acts

[Quicksaver]

Problem

For reviews, plans and design questions I want the answer challenged by models from different providers, not just one. Orchestrator V2 already lets the agent delegate the same question to several providers and follow up with them, so I can ask for this in a prompt. But each time the agent improvises the rounds, keeps its own tally of who agreed with what, and suggestions raised along the way get lost between rounds. There's no setting for how much agreement is enough, and no record of the outcome beyond the transcript. Related: the adversarial-dialogue comment in #6911 and the independent review stage in #6818.

Proposal

Magi, a review panel owned by one conversation:

• Panel: 2 to 9 participants. Each is a model on any provider, with an optional personality and a weight from 1 to 100. You also set a consensus threshold and a round limit (default 1).
• Personalities: each participant can take a perspective from a catalogue in Settings → Magi. Included ones are security, reliability and operations, maintainability, accessibility, product and UX, a skeptical reviewer, and more. You can edit them or add your own. Each run keeps the personalities it started with.
• Starting: set up the panel and arm your next message, or ask the agent to use Magi. The agent never starts Magi unless you asked.
• Rounds: participants are ordinary child conversations created through delegate_task. They answer the first round without seeing each other. In later rounds, the same child conversations see each other's latest answers and vote on the agent's proposed conclusion and on every suggested change.
• Arbitration: the conversation's agent classifies whether each participant supports its conclusion. The server adds up the weights and decides whether the threshold is met. A participant that fails or abstains counts as zero but stays in the total. The agent reports the outcome along with any remaining dissent.
• Acting: only the conversation's agent acts. Participants are told to stay read-only and can't start Magi themselves. An enforced read-only mode for delegated tasks, like the one proposed in Read-only runtime mode for delegated tasks (Codex: read-only sandbox, approval never) #15184, would strengthen this later; Magi doesn't depend on it.
• Evidence: the agent hands participants a finished tool result, such as a diff, by its id. Each participant reads the full copy instead of the agent pasting it into every prompt.
• Visibility: each run shows in a Magi right-panel tab and as a summary in the conversation, with token totals. On mobile it's a sheet in the conversation menu. The agent can pause, resume and stop a run.

Interactive demo: https://quicksaver-piz.vercel.app/t3code/magi-consensus-orchestration

Why not just a skill

A skill on top of delegate_task and t3_thread_send could prompt similar rounds, but the outcome would depend on the agent following the protocol every time. Magi makes the orchestration deterministic. The server dispatches every round to every participant, enforces the round limit, computes the weighted result, and keeps the record of answers, votes and suggested changes. The agent can't skip a participant, miscount, or lose track between rounds, and every run can be reviewed afterwards. Magi also covers what a prompt can't: arming a run from the UI without writing the instructions, sharing a full tool result with every participant without pasting it, and run history on web and mobile.

Suggested slices

The current implementation is large (about 115 files, built on V2), so I'd split it:

1. Server and agent tools: runs, rounds, the weighted threshold, pause/stop, and sharing tool results with participants as evidence, started by the agent on request. No new UI.
2. Web: the panel to configure and arm a run, run history, the summary in the conversation, and a settings page for personalities and arbitrator instructions.
3. Mobile: the same, in a sheet.

Decisions I'd like maintainer sign-off on

First, a yes or no on the proposal as a whole. Beyond that:

1. Surface: its own right-panel tab and settings page, or runs shown only in the conversation?
2. Cost: rounds default to 1 and unlimited rounds show a warning, but participants' own subagents aren't capped. Is that enough, or do you want a hard cap?

Out of scope

Automatic model routing for delegation (#15182), sequential pipelines (#6818), and splitting work across parallel agents (already covered by delegate_task).

Contribution

• I would be open to helping implement this.

If the direction works for you, I'll open the first slice scoped to whatever we agree on.