Arena

Arena

Fan out N parallel attempts at the same task. Read every candidate end to end. Pick the strongest as the base. Graft the best ideas from the others into it. Verify the synthesized result.

Start

Open a todolist with one entry per phase before launching anything. The arena runs autonomously and the list keeps phases from silently disappearing.

1. Frame
2. Fan out
3. Cross-judge
4. Pick
5. Graft
6. Verify

Phase A: Frame

The N candidates will receive the same prompt, so the prompt is the contract. Get it right before spawning anything.

1. State the artifact each candidate is producing.
2. Derive the rubric. State what success looks like for this task, then turn it into 3-6 concrete gradeable criteria. Concrete: Adds a --dry-run flag that skips writes. Vague: code is correct. The rubric is the picker's tool in Phase D; candidates only see the task.
3. Pick the runners. Default 4: claude-opus-4-8-thinking-xhigh, gpt-5.3-codex-high-fast, gpt-5.5-high-fast, and composer-2.5-fast. Spawn more when the arena covers multiple design directions. Same model N times when the work is generation-bound rather than judgment-sensitive.
4. Assign output paths. Each candidate writes to its own location (a git worktree where possible, otherwise /tmp/arena-<slug>/candidate-<n>/). N candidates writing to the same path is shared mutable state and fails the the separate-before-serializing-shared-state principle skill test.

Phase B: Fan out

Spawn all N subagents in one message with run_in_background: true, each with the task, the path to the shared grounding, its own output path, and instructions to produce both the artifact and a short rationale.

The rationale is mandatory. Without it, the parent cannot tell whether a candidate's structure is principled or accidental, which makes Phase E grafting unreliable. Each rationale names the alternatives the candidate considered and what it rejected.

If a candidate fails to produce output, proceed with N-1 and note the dropout in the synthesis record.

Phase C: Cross-judge

After all Phase B candidates complete, spawn one readonly judge subagent on a different model family from the parent's. It sees the rubric and the candidates by path label, scores each criterion, and recommends a base with rationale. It runs in parallel with the parent's reading in Phase D, not with the candidates themselves. Spawning while candidates are still writing means the judge sees partial or empty outputs and reports them as dropouts.

Phase D: Pick a base

Read every candidate end to end before picking. Skimming N candidates surfaces only the candidate whose surface looks most familiar.