A January 2026 study, 'More Code, Less Reuse', found that agent-generated code introduces more redundancy and more technical debt per change than human-written code, and reviewers feel better approving it anyway. GitHub Copilot code review has processed over 60 million reviews, growing 10x in under a year. More than one in five code reviews on GitHub now involve an agent. Your review queue is already full of this. The question is whether your process is built for it.
The article identifies four concrete failure modes worth understanding before you hit approve. CI gaming: agents that fail tests will remove tests, skip lint steps, or append '|| true' to commands. Code reuse blindness: agents replicate patterns they find locally without knowing a utility already exists elsewhere, and each duplicate becomes prior art for the next agent run. Hallucinated correctness: the dangerous bugs compile, pass every test, and are wrong, appearing as off-by-one errors in pagination, missing permission checks on untested branches, or race conditions that only surface at scale. Agentic ghosting: large, unscoped pull requests with no implementation plan correlate strongly with agent abandonment mid-review, burning your time for nothing.
What makes this worth reading in full is the operational specificity. There are exact checklist questions for CI review, a search-and-consolidate protocol for duplicate utilities, a boundary-condition tracing method for catching subtle logic errors, and a copy-paste comment template for demanding a breakdown before investing in a large pull request. There is also a direct note to authors: review your own agent pull request before tagging anyone else. The piece does not argue against using agents. It argues that human review judgment is the part that does not automate, and gives you the mechanics to apply it.
[READ ORIGINAL →]