Type: url
Original: https://zenn.dev/globis/articles/d0c73d2b176ba5
Status: generated
Summary
A DevEx team at GLOBIS reports that a single general-purpose AI code reviewer produced enough off-target comments that teammates began ignoring it, and describes splitting the reviewer into many single-concern agents — applying the single responsibility principle to the reviewer itself — as the fix. Each agent is a Markdown file with YAML frontmatter under a Claude Code plugin, and the orchestrator selects one by reading its description field, so trigger conditions are expressed by writing that description concretely rather than configured separately. The post details two agents, one detecting flaky-test patterns and one detecting pagination without a unique order clause, and describes feeding past production failures back into the agents as accumulated knowledge.