arena

Use the arena workflow: frame a shared prompt with a rubric, spawn N subagents in one message with run_in_background set to true, and assign each its own output path. Each candidate produces the artifact plus a rationale for later comparison.

Use it for non-trivial artifacts where one attempt would lock in the wrong design shape, such as architecture decisions or complex modules. Skip it for simple, well-understood tasks where parallel attempts add no signal.

A cross-judge subagent, preferably on a different model, scores each candidate against the rubric while the parent reads every candidate end to end. The base is picked on which candidate a maintainer can extend most easily, with agreement between judge and parent confirming the pick.

The workflow proceeds with N-1 candidates and records the dropout in the synthesis note. The remaining candidates are still judged, grafted, and verified normally.

If candidates wildly diverge, the task framing was under-specified and the arena should be reframed and re-run rather than averaged. The synthesized result also still requires verification, since the arena does not exempt output from normal scrutiny.