People often have an assumption that productivity gains from AI coding agents are going to come from generating code, which in turn will allow us to complete more tasks. I think that this is the wrong approach to sustainable productivity gains with AI in enterprise environments.
First of all, coding speed is rarely the bottleneck in many enterprise environments. If we imagine that the time taken to implement the code for a given task goes to zero, the bottleneck would immediately become quality assurance (QA) and code review. This of course assumes that the task is well specified upfront, and that other teams work are not blocking the changes.
I rarely hear developers say that code review is the favorite part of their work, and it's the same with QA. Yet we spend our focus on generating code with AI agents that we have to sift through and review, before giving our poor coworkers the same "joy" of reviewing soulless AI-generated code.
Shouldn't our focus rather be on using AI agents to improve our code review and QA experience, so we are left with the fun part - building?
For example - my favorite way of going through pull requests is to sit together with the PR author and talk through the changes. I usually have a lot of questions, and it's much faster to get answers in real time than going back and forth in comments. The agent can answer ~90% of my questions and even prototype alternative solutions. That means I show up to the review with fewer, but much higher-quality questions. It's a much better experience for both sides.
I used to focus on using AI to generate code. Now I think the bigger opportunity is using AI to improve the thinking around the code-review, validation, and decision-making.
Are we over-optimizing for code generation and ignoring the real bottlenecks?
Top comments (0)