Just today I reviewed a patch made by a teammate using Claude for a not very significant bug, where in addition to fixing it (in a very roundabout way), it introduced 2 new very subtle bugs, and refactored a large portion of another function that was unrelated to the code, and made several observations and comments which were subtly wrong and out of touch with the rest of the code.

I ended up rejecting the PR and writing my own very small local patch for the bug.

Other people in the team use Claude to help them review my own PRs, and it sometimes finds minor issues, although there was only a single time it found a major issue.

My own use of LLMs is mostly to develop large features, which I then thoroughly review myself.

I'm concerned that there are people pushing on the code generation front, the code reviewing front, and the laziness front. Eventually they're going to all meet in the middle, and there will be a significant number of engineers who are using AI to write and review the code, rubber stamp it, and push it into prod, with disasterous results.

Day Job: Similar but more human-in-the-loop and we are still feeling out "what needs human review" and "agent review is good enough".

LLM reviews have shown me/us that they will catch more things (legit things and silly things, to be fair) than human reviewers. Human reviewers are very bad at seeing what's _not_ there. Yes, humans will catch some things that the LLM won't but it's normally only on very custom things we've done that the LLM isn't "trained" on. [0]

I think that as time goes on code reviews will be almost fully automated and humans will focus more on the plans before building, overviews of what was built, and other spec/diagram-type "artifacts" than looking at the actual code.

[0] For my side projects and even a little for work I've been trending in the "do it the way the LLM wants" not "force it into the shape you thought of" because it's easier for the LLM to write/maintain/"understand". Similar to how I don't tell an LLM "put a red button on the page here that does X", I give it the problem I'm trying to solve so that I don't "bias" it to the way I first thought of. I may force it to do it "my way" in the end but I find I get exposed to new ideas or new/different ways to solve my problem when I don't "lead the witness".

Here’s the interesting part: each PR gets reviewed by more than 3 people, each from their own perspective. Based on the questions and the affected paths, the reviewer directs the agent to focus on different points. Someone who is already familiar with the codebase can also guide the agent in the right direction.

Recently, we also started automating UI reviews. The agent runs all the services, verifies the feature through the UI, takes screenshots, and uploads them directly to the GitHub PR. We call this UI Review. + Test coverage also improtant.

How does this work exactly -this is your current workflow? Which tools are you using . I assume you mean your human reviewers are using ai tools interactively and deciding which fixes the tools propose will be implemented ?

The speed you're as able to move at with no to lightweight manual code reviews is just so much higher, but there are cases where broken code could mess up critical prod systems so I don't see how you'd get rid of human code reviews there just yet.

For personal usage I've found quick DeepSeek flash reviews a massive boon to finding obvious bugs though.

But in cases where context lives in peoples heads, its hard for ai to know whats going on snd whats the main goal of this code chanhe.

if the context is given (or documented somewhere claude has access to) + context of company is not too big ti consume it, there is no use in human review imo.

https://github.com/jacquardlabs/gauntlet

If something highly sensitive, yes, and in general, a tendency to review with more scrutiny on backend changes.

But for frontend, less and less every week.