arena

Use the arena workflow: frame a shared prompt with a concrete rubric, then spawn N subagents in one message with run_in_background enabled. Each candidate writes to its own output path and produces both the artifact and a rationale.

Score each candidate criterion by criterion against a rubric derived from the task, and compare your scores with an independent cross-judge subagent. Pick the base a future maintainer can extend most easily, then graft the best ideas from the losers.

Use it for non-trivial artifacts where one attempt would lock in the wrong shape, such as ambiguous design or architecture tasks. Skip it for simple, well-specified work where the overhead of N candidates adds no signal.

The workflow proceeds with the remaining N-1 candidates and records the dropout in the synthesis note. The final record names the base, grafts, rejections, dropouts, and the verification result.

Each candidate writes to its own git worktree or a unique /tmp directory so no shared mutable state exists between them. N candidates writing to the same path would corrupt outputs and violate the separate-before-serializing-shared-state principle.