Do We Know What We're Doing?

It’s easy to convince yourself that you’re moving fast when using AI. You can “write” (generate) a plausible looking design doc within minutes. An agent can close all of the tasks on your sprint board overnight. But sometimes when you take a step back from all of your agent sessions and inspect the results, you realize all of those tokens have gone to waste.

Let’s think about pull requests for a moment. Merging one pull request per day used to be a good gauge for productive software development. A well-scoped PR was a self-contained commit with no more than 500 lines in the diff. This is small enough to where the author can make incremental commits at a consistent pace and the code reviewer can understand the change in full. Recently, I’ve been tagged on several multi-thousand line pull requests in a single day. Behemoth pull requests are a chore, but they’re only a minor symptom of a much bigger problem. In the last few months I’ve experienced:

• Asking why a function was added to a pull request and the answer is “I don’t know”
• Leaving a design review where no one, including the “author” (meat proxy), knows what they’re trying to propose
• Learning that the team is hustling to release an API that doesn’t actually have a clear use case
• Being reminded that I forgot I already implemented something weeks ago in a blur of multi-tasking

So my question is, do we know what we’re doing? We’ve entered the squiggle of uncertainty but can’t seem to find the path out.

I think it’s still important to understand what we’re building. Understanding the code means that I can tell fact from fiction when working with an agent to investigate a bug. Understanding the system means that I can predict when a component will reach a scaling cliff before it happens. Understanding the product means that I understand when to NOT ship a feature because it adds significant complexity without much benefit.

Everyone wants to move quickly. If you don’t know where you’re going, you might make it far, but you won’t end up at the destination.