Agent Roster

Run a named roster through independent first positions, evidence-based deliberation, explicit objection resolution, and a late consensus check.

Parse the request

Prefer a short form:

<model>, <model>, ... | <task>

Examples:

sol, terra, luna | decide: should we adopt this architecture?
opus x2, sonnet | improve: proposal.md

Natural-language rosters work too. Optional controls can follow the task: judge=..., lead=..., consensus=..., or budget=....

Preserve requested model order and repeated seats. Resolve names against the harness's actual model controls. Never silently substitute, omit, or add a requested model. If the harness cannot honor the roster, report the limitation before running. If the user gives no roster, choose three meaningfully different available models or variants and announce them.

Set the consensus rule

Honor an explicit rule such as unanimity, supermajority, judge decision, or verifier threshold. Otherwise use qualified consensus: every seat must return accept or accept-with-caveat, and no unresolved objection may identify a material factual, logical, safety, or verification failure. A material objection must identify the affected issue, explain its likely effect, offer checkable support, and state a resolution path. If that bar is not met, report no consensus rather than manufacturing one.

Run the roster

1. Lock a shared brief: objective, inputs, constraints, deliverable, evaluation rubric, consensus rule, and comparable per-seat budget.
2. Give each seat the brief without peer positions. Collect its proposed answer or artifact, evidence, assumptions, confidence, and likely failure modes.
3. Build a comparison matrix of distinct claims, agreements, conflicts, evidence, and unknowns. Do not treat vote count as proof.
4. Cross-examine the disagreements. Send agents the conflicting claims and evidence, not persuasive peer transcripts. Ask for counterexamples, falsifying tests, missing evidence, or concrete repairs.
5. Have every seat revise independently and state what changed its view, what remains unresolved, and its current confidence.
6. Draft the consensus result only after that revision. Circulate it late for accept, accept-with-caveat, or reject, each with a specific reason.
7. Run available tests, source checks, or held-out review. Verifiable failures override votes. Preserve dissent and stop at no consensus when a material objection survives.

Before cross-examination, agents may share scores, failing tests, counterexamples, verified sources, interface contracts, and blockers. Do not broadcast complete peer solutions or persuasive summaries. The lead routes communication; avoid unstructured all-to-all debate.

Route by task

• For the full consensus contract, message schema, and shared-workspace rules, read references/protocol.md.
• For new or existing research articles, read references/research-writing.md.
• To explain or modify the design, read references/interaction-tax.md.

Report

State the requested and resolved roster, consensus rule, comparison, decisive evidence, objections and their disposition, final verification status, and any minority report.