Claude Code gated workflow re-reviews consume 75% tokens

An audit-trail study of coding-agent workflows using Claude Code revealed that re-reviews consumed over 75 percent of the reviewer subagent's tokens, with nearly half following reports that had found no blocking issues. The author argues that review loops must be capped and stop rules moved outside the model into deterministic scripts to prevent excessive token waste and unreliable self-judging orchestrators. Open article →