At OpenAI, I build agents. At home, I keep finding jobs for them.

There has never been a better time to be a builder.

A fuzzy idea can become a working thing in an afternoon. The odd little problems that are too personal, too niche, or too annoying for normal software are suddenly worth solving.

That is what I do here.

I like software that starts with the way a person actually thinks:

I have a folder of medical scans. Can you help me understand them?

Is the pool ready?

Did the gate actually close?

Then I work backwards until an agent can help: clear state, useful tools, honest answers, and a person still in charge.

Agent-first software sounds fancy. Mostly it means I should be able to say what I want, see what the system knows, and trust it to admit when it doesn't.

Point it at a folder of MRI or CT scans. An agent finds the study, opens a useful view, moves through exact slices, points at what it is talking about, and helps you ask better questions. The scans stay on your computer.

Medical imaging is complicated. Getting oriented in your own scan folder should not be.

Annoying apps are an excellent source of projects.

• gatectl — gates and garage doors
• poolctl — pools, pumps, and cleaners
• hottubctl — hot tubs and temperature
• switchctl — switches without the mystery

Same idea every time: show the state, make the useful action easy, slow down the consequential ones, and check the world afterward.

The *ctl tools are one small corner of this. They are also a pretty good explanation of how I think.

Agents that can do useful work. Systems that explain themselves. Local software that respects the machine it runs on. Tools that cross the line between bits and the real world.

Mostly: powerful things made simple enough to use.