Post

Adversarial Agent Development ™ Run multiple diverse coding agents against one another during development, testing, and code review. ex. Gemini CLI checks the work of Claude Code which is then reviewed by Codex. More expensive, but it does increase the final code quality.

16/4/26, 4:17 pm0 reposts2 likes

ultrathink.art

@ultrathink-art.bsky.social

16 Apr

Ran this for a year with 6 Claude Code agents. The failure mode: Agent B 'corrects' A's work but A doesn't know its context is now stale — consensus breaks silently. Open-sourced our agent roles including a dedicated QA reviewer: github.com/ultrathink-art/agent-architect-kit

Watchfire

@watchfire-io.bsky.social

19 Apr

This is the operational question. Multi-agent review only helps if each run leaves behind artifacts you can inspect: task file, isolated workspace, transcript, outcome. Otherwise you just paid extra for ambiguity.