Arena

Arena never grants authority beyond the caller's task. Candidate agents may make only the local changes the user already authorized. They must never publish, push, comment, review, message, or mutate an external system. Subagents may communicate with each other, but no agent may communicate with a person. Repeat that sentence verbatim in every child prompt.

Model and reasoning preferences are optional. Read repository-documented model configuration or .agents/agent-models.md when present. Use the caller's confirmed candidate choices when supplied; otherwise use design candidates. Select one entry from independent judge for the cross-judge. A candidate list requests one seat per entry, including aliases. The judge list is a pool, not a panel.

Read its optional reasoning profile line and validate each explicit model/profile pair against the current host. For a missing role or an auto or inherit-parent model entry, omit both overrides. For a missing, auto, or inherit-parent profile, omit the reasoning override. Never require the file, construct model identifiers from effort labels, or use unconfirmed choices. If a configured choice is unavailable or the host rejects a model/profile pair, report the fallback and retry that seat once with both overrides omitted. If the retry fails, drop that seat and say so. If dropped seats leave fewer than two candidates, or one when only one was requested, add candidates until that count is met. Run each added candidate on a remaining candidate's model, or with both overrides omitted when none remain, and say so. If the host cannot run any seat, use the local fallback below. When engineering-mode is installed, match each subagent's reasoning profile to the job by its Delegation rule; otherwise use the configured profile.

If the host cannot run parallel subagents, produce the candidate attempts sequentially in isolated local paths. If isolation is also unavailable, keep candidates as clearly separated drafts and do not let one candidate see another before it is complete.

1. Frame

State:

• the exact artifact each candidate must produce;
• the shared grounding and scope;
• three to six concrete, gradeable criteria;
• the number of candidates;
• the selection rule: first passing result, rank all, or best overall.