Post

I was pairing with a dev who commented on the fact that I always resolved merge conflicts by hand versus using a UI. I joked that it's because I'm a control freak and without missing a beat responded "Yup." 😅 Let me tell you about my experiences with agentic coding these last 3 months...

So this role I'm in now was a tiny 2 person team. Two senior ICs managing a family (50~) of APIs and libraries for that last decade. These services are providing a public interface for data produced from various sources (integrations, dev teams, jobs) back out to dev teams to work with.

Those two senior level ICs both retired at the same time. They were an all knowing, doing a lot of "hidden" work to keep things going and it was a complete silo. When I applied for the role, I missed the part about it being an AI "augmented" role. I noticed an hour before first the interview.

At this point I had used LLMs to explore implementation ideas at a small scale. My first attempt implementing Ackermann steering for a video game. It was a single function that I reviewed and integrated into the game. It worked.

That's how my AI uses with coding went for a few months up to, or to explore why code I wrote didn't work the way I expected. I didn't trust anything it said but it would atleast give me starting points.

I rejected code at a 20:1 ratio easily. It didn't feel like a performance gain. The bigger the task, the worse the rejection rate. It was frustrated as context slowly gets polluted and the problem got worse.

I was clear to me any AI usage during development was best, for me, when it's tiny and focused. When it was drudge work, not design work. For me telling it how to build worked better than what to build. I still feel this way...