AI coding agents now write code faster than humans can review it, and the data says that’s starting to hurt. As Hacker News reports, a 2026 study by Faros looked at two years of telemetry from 22,000 developers across 4,000 teams. In organizations with high AI adoption, engineering throughput went up. Bugs, rework, review load and production incidents went up even faster.
The numbers are hard to ignore. When Faros compared periods of lower and higher AI adoption, it found:
- 54% more bugs per developer
- About 5x longer median PR review time
- More than 3x as many incidents per PR
The article is careful here, and you should be too. These are correlations, not proof that AI caused every bad outcome. A later Q3 report found some of the early effects were easing. But it also found that verification, remediation and operational costs kept piling up downstream.
🔍 Why This Matters Now
The bottleneck has moved. Writing code used to be the slow part. Now it’s checking code. A small team can’t meaningfully review huge AI-generated diffs, and normal review habits are already falling behind.
The article’s answer is automated software verification. The idea is to use the same agents, skills and harnesses that speed up coding to build testing and review automation. It doesn’t promise bug-free software. It promises more chances to catch problems before users do.
What stands out to me is that this flips the usual AI pitch. The selling point stops being “write more code.” It becomes “trust the code you already have.”
🛠️ The Tools Competing for the Verification Layer
The piece walks through what’s available right now. It splits into two camps.
Methodology and local-loop tools:
- Superpowers. An agentic development method and skill collection that pushes agents to act like disciplined engineers. It covers planning, test-driven development, debugging and reviews by separate agents. The tradeoff, per the article, is that running the full method can slow you down on some tasks.
- RoboRev. Built by Wes McKinney, the creator of pandas. A local daemon reviews every commit with a configured coding agent, so you can fix issues while the context is still fresh. Using a different model as reviewer gives you a second perspective, though the article notes that’s “not a guarantee of an unbiased or correct review.”
Hosted PR reviewers:
- Claude Code Review. Anthropic sends several specialized agents to dig into a pull request in parallel, hunting for logic errors, security holes, regressions and missed edge cases. It’s a research preview for Team and Enterprise plans with separate usage-based billing. There’s also a lighter /code-review plugin for Claude Code, and the article warns not to treat the two as the same thing.
- OpenAI Codex Review. Trigger it with @codex review on a GitHub PR, set up automatic reviews, or check local changes before a PR exists. It supports a review-and-repair loop, though the fixes still need independent checks.
- GitHub Copilot Code Review. The most hands-off option. It auto-reviews PRs and suggests fixes. The risk is that teams start accepting its feedback without thinking.
Greptile, CodeRabbit, Qodo and Cursor are also in this space. That’s a crowded market, and it tells you where vendors think the next fight is.
📈 The Bigger Picture
Every major model lab now sells both the code generator and the code checker. That’s good business. If AI creates a review problem, AI review becomes the upsell.
It also raises a fair question. Can a model reliably catch mistakes from the same model family? Using a different model as reviewer is one hedge. Old-school checks like tests, static analysis and security scans are another. The article keeps coming back to the same point: no single plugin or reviewer makes code bug-free or vulnerability-free.
✅ What Teams Should Do
- Measure first. Track bugs, review time and incidents per PR before and after you scale AI use. Faros-style telemetry is now table stakes.
- Move review earlier. Catch problems at commit time, not at the end of a sprint.
- Mix your reviewers. Pair AI review with tests, CI checks and a human who owns the decision.
- Never auto-accept. AI review comments are suggestions, not verdicts.
- Check the bill. Deep multi-agent reviews can cost real money. Read your plan terms.
My recommendation: treat verification as its own engineering investment, budgeted and staffed like the coding agents themselves. Teams that only scale generation will drown in review debt. Teams that scale verification alongside it will keep the speed without the 3 a.m. incidents. Full details on each tool are in the original piece on Hacker News.