Same task, same prompt, Claude Code and OpenCode in separate worktrees. An orchestrator picks the winner and steals the best parts from the losers. Sometimes the winner is Qwen3.8-27B on my local server, not Opus.

After each task they "dream": they consolidate what they learned into skills and agents, pushed to a repo so every project gets it.

It burns more tokens, but I review a lot less. Full write-up here:

https://menetray.com/en/blog/how-i-work-ai-today-ddev-worktrees-and-arena-where-my-agents-fight-it-out

Anyone else running more than one harness on the same task?