AI Amplifies What's Already There

I went to Tropical Rails, the conference of the Ruby and Rails community (a programming language and its framework for building web apps), with the usual expectations: see what the community is up to, check out new tools, talk to people. I didn’t expect the entire conference to orbit around a single topic, but it did: AI and agents, meaning programs that use an AI model to carry out tasks end to end, not just answer questions. Not in the “check out this new tool” sense, but in the “this changes the rules and we need to talk seriously” sense.

The takeaway

What I walked away with, and the thread running through almost every talk, is this: language models (LLMs, the kind of AI that writes text and code) and agents don’t have opinions about code quality. They amplify what’s already there. If good patterns go in, good code comes out. If tech debt goes in (the shortcuts and patches left in the code that someone has to pay for eventually), more tech debt comes out, and faster.

If the codebase, meaning all the code in the project, isn’t clean and ready to grow, it won’t be able to keep up with the product. And agents now produce code at a pace no human can review. With that, tech debt stopped being just a quality problem and became a capacity problem, a matter of how far the team can get.

This was true before too, but it never mattered this much.

Two tools that stuck with me

Two gems (that’s what Ruby calls packages of ready-to-use code) came up multiple times in the talks and are worth mentioning.

evil-seeds is like a database dump (a copy of everything that’s stored) but smart. You can configure how the data gets scrubbed so nothing sensitive is left in, which tables to include, and how many records to pull. Sounds like a good path to finally having a clean, safe copy of production (the real database, the one customers use) to work with on your own machine.

Blazer is for spotting weird or inconsistent data in production. The idea of having something that keeps watch over data health all the time is something we could use, whether or not we adopt this specific tool.

The real takeaway

If we’re serious about using AI to move faster, the codebase needs to be in a state that makes that possible. Agents amplify tech debt, they don’t fix it. I came back from the conference wanting to pay down technical debt, which is not something you hear people say very often.