Skip to main content

Agents Write More Code in a Day Than I Did in a Week

Updated 2 min read By Matt Bartlett

TL;DR

AI agents can now generate, review and test code, but the person directing them gains output faster than understanding. The result can feel like inheriting someone else’s codebase - except that someone was you.

AI has changed the pace of software delivery in ways I’m still getting my head around: generate via agent, review via agent, test via agent. The spec gets satisfied. The tests pass. The output appears to work.

What I recognise now though is that I’m accumulating output faster than I’m accumulating understanding.

What changes when you build it yourself?

When I build something myself, I build a map in my head. I have a strong idea of where that edge case may live. I know why I wrote in that particular condition. I remember the scenario this function solves.

After an AI-assisted sprint I have working software that I didn’t fully author, and that gap widens every time I hand off more work - every time I use an agent to solve a problem I’d previously have solved myself.

What happens when you inherit your own codebase?

So when something complicated breaks that the agent struggles to fix, I dig in to help and realise I don’t have the same instincts of where to look that I used to build naturally.

It feels like inheriting someone else’s codebase - only that someone was me.

How do we mitigate this? Include “document your code thoroughly” in prompts? I’d love to hear suggestions.

See the original post on Matt's website: tiltedsky.net(opens in new tab)