A lot of our code is now written by AI agents. A few years ago that would have sounded reckless. Today it is part of how a small company can make more than one product.
What agents are good at
An agent is fast and does not get bored. Give it a clear task and a way to check its own work, and it finishes sooner than we would. It also does the parts people skip, such as tests and small fixes.
That changes what is worth doing. Polish we would once have cut for time now gets done.
Where people still do the work
An agent builds the wrong thing as fast as the right thing, and sounds equally sure about both.
So deciding stays with us. We choose the problem, we decide what the product will not do, and we judge whether a screen is clear or only looks finished. We read what gets built and we use it. We say no often.
Products that agents can use
More people will use software through an agent of their own. A product that only works through its screens leaves them out.
We want each of our products to work both ways, for a person clicking and for an agent acting for that person. Some of ours already do. We expect this to be normal soon, so we are building for it now.
What we are still learning
We have not figured this out. We still misjudge how much to hand over and how closely to check the result. We will write down what we learn here.
