If you hand a new project to a human, they will start out pretty useless. Over time, they will develop experience, understanding, and intuitions. Their ability to reason about it, answer deep questions, and do meaningful work will continue to develop to the point where they become an expert (assuming they’re motivated).

An AI agent is the opposite. It’s like the Benjamin Button version of the novice → expert timeline. When the project is small, it presents like the expert. It can quickly analyze, answer questions, and do real work. As the project grows, the fundamental limits of it’s context window start to show through. It increasingly engages in hackery; punching holes through abstractions, duplicating code, deviating from guidelines, and making changes without understanding the implications.

Your trust in a human working on this project would grow over time. Your trust in the agent recedes.

A mistake I keep seeing (and making) is positioning the AI above the human. It’s a pretty good advisor but a terrible lead. Any time I start to feel the pull towards “sure agent; whatever you think sounds good” I know it’s time to power it off. Good judgement requires understanding, and it’s dangerous to let that understanding slip. Amazon’s requirement that seniors review all code changes made with AI assistance is evidence of how the work must be coupled with an expert who can hold the big picture in their head.

Interestingly, the AI seems to pave both paths. It’s easier to offload your thinking to the machine and easier to pursue a deeper understanding. Being able to quickly gather and synthesize information and drill into any part of a concept that feels fuzzy is truly incredible. At the same time, you’re a keystroke away from “getting the thing done” in a way that will probably work fine, for now.

I think we’re still in the formative period. It’s really only been about 6 months since “agentic coding” has been acknowledged to be useful by anyone other than the people selling it. I’m trying to make sure my mental model shapes up correctly, and to the degree possible, that my habits and decisions are healthy and human-centric. I think emphasizing the assistance in AI assisted coding is vital.