As AI agents spit out tokens at even faster speeds, it’s increasingly obvious that code review is now the bottleneck with development. I guess, in some abstract sense, code review was always the bottleneck with development teams and a few obvious solutions existed for streamlining the process:

• Clear coding guidelines so that a contributor can judge whether their code is compliant with standards expected by the team. As an experienced developer, there are always some unwritten rules, but I’ve never seen a team regret explicitly writing this down, even if it’s a brief and concise list.
• Write good tests. It is always worth the effort. (Good) Tests easily communicate your expectations and blindspots to other developers.
• When in doubt, look at existing code and documentation for figuring out how things should be done. When you’re finished, leave the code and documentation in a better state than you found it.
• Plan and discuss major decisions before you jump into the implementation. Get a general consensus among team members (even if it means “disagree but commit”) and decide metrics to guide future architectural modifications.
• Use automated tools for the boring stuff (formatting, linting, etc) as much as possible.
• Have your merge request reviewed by one or more developers for a final seal of approval.

I’m probably spending the most time on the last point as an AI agent orchestrator, and it sucks. Reviewing code line-by-line is tedious and there’s no clear reward at the end of it (which is why in the real world, people either nitpick too much or just go “Looks good, ship it!🚀”).

AI agents are still pretty crappy at all of the points above though, but clearly getting better over time. I’m also getting better at reviewing more code than I was previously but unsure how I can improve my endurance and effectiveness on this task (can you seriously review 2-10k LoC everyday?).

I know there’s a few developers who are admitting on X that they’re shipping code without review as the agents are so good. This is a fine approach for anything fun, but quite dangerous for something even half serious. It’s a race against time though as agents continue to improve and we might see particular models which are better reviewers than coders.

As of right now, I’m not taking the “just throw more agents at it” approach. I’ll continue experimenting to improve my personal, human workflow to review agentic output and report back.