A paper from KTH Royal Institute of Technology argues that human code review — the primary quality gate in software development since Fagan formalized code inspection in 1976 — is now superseded by LLM-based coding agents. Not supplemented. Superseded.
Author Martin Monperrus frames the argument around what coding agents can already do:
- Read and understand code across full repositories, not just the diff a reviewer sees in a PR
- Write, test, and repair software autonomously — closing the loop that human reviewers can only flag
- Operate continuously without the bottleneck of waiting for a colleague’s calendar to open up
- Apply consistent standards across every change, eliminating the variance between a careful reviewer and a rubber-stamp approval
The core claim: if an agent can generate code, test it, fix failures, and verify the result against specifications, the traditional review step — a human reading diffs and leaving comments — adds friction without proportional value.
This is a position paper, not an empirical study. Monperrus is making a deliberate provocation grounded in where agent capabilities are heading, not reporting experimental results that prove human review is worthless today. The distinction matters. Practitioners like Sean Goedecke have argued the opposite direction — that code review skills are exactly what makes someone effective at using agents, because LLMs still lack the depth of judgment a competent engineer brings.
NOTE
The paper traces code review’s lineage back to Fagan’s 1976 inspection process. Five decades of practice is a lot of institutional weight. Even if agents handle correctness checks better, review also serves knowledge transfer, mentorship, and shared ownership — functions the paper doesn’t fully address.
The real takeaway isn’t “stop reviewing code.” It’s that the ratio of human-to-agent review effort is shifting fast. Teams already using agents for automated PR checks — tools like Bito’s CodeReviewAgent or parallel Claude Code sub-agents — are running the experiment in production. The paper gives them an academic frame for what they’re already seeing: agents catch more, faster, and don’t get tired on Friday afternoon.