Code reviews are important
They are also boring and time-consuming. This point of view will resonate with many senior engineers and managers.
When I first started working somewhere with mandatory code reviews before merge it was a revelation. We have been talking about doing this for years, I thought, and here we are doing it on every pull request. I am in no doubt that it improves code quality and spots bugs before they become issues. The challenge is they can block development flow, both of the code being developed and also of the code the reviewer would otherwise have been working on. I have seen code sitting for weeks awaiting review and also features not shipping on time as the developer had to go and review another person's code. Neither of these are great and it's a tough choice what developers should focus on.
Fortunately, we are now in an era where there is an alternative choice. We can get AI agents to review code. In 2025 the frontier models got good enough at writing code to be the primary developers in some forward-looking organisations. By early 2026 the open-weight models could do the same. It's not always perfect but then neither is any human engineer. If AI agents can write code as well as a mid-level engineer then it's reasonable that they can review code just as well. With the right prompts it is likely that they can review code more like a senior engineer. They may not have the full context that a principal has but even that can be fixed with the right prompts. The great thing is that the agentic code reviewer sits on its own and so does not block other development and it can run in minutes on every single code review.
Focused reviews
AI agents are best when they are given clear, specific instructions. Contextual guidance on what they should focus on will help them hone in on problems and present clear findings back to you, the operator. Agents can be asked to think like a performance engineer or an architect or conduct a general code review with specific coding standards. So we run multiple agents each with a focussed prompt. We can call on others who have published prompts for specific use cases, as is the spirit of open source.
Run it every time
In that spirit, there is a Github action runner for automated code reviews available here github.com/nickwhiteley/pr-reviewer, which uses a config file in your repo to direct the PR reviewer.
None of this stops human reviews and these are definitly important to still do but automating reviews reduces the pressure and often catches things that people miss.