mw/toolkit
A small corner of the internet for the work I care about: the systems that think, and the ones that move. This is where I keep my notes, tools, and whatever I am building at the moment.
I can't stop building agents.
I run agents every day. I have them wired into almost everything I do: coding, debugging, looking up things, kicking off long jobs while I go make coffee. The thing that hooked me wasn't the model. It was watching the loop come together: a plan, a tool call, a mistake, a correction, a working answer, and somehow a sentence at the end that wasn't nonsense.
I keep coming back to the parts of this field that aren't the leaderboard. Who's actually shipping? Whose agent is the user in control of, and whose is the user just watching? When the answer is wrong, who fixes it, and how fast? Those questions are more interesting to me than another 5% on a benchmark.
The other thing I spend time on is LLM proxy fraud: the slow leak of resellers rebranding cheap backends as premium ones and selling access as if it were the real thing. It's the kind of work nobody brags about, but the trust the whole agent economy runs on sits on top of it, and I'd rather know which model I'm actually talking to.
Then there's the part with motors.
A model can hallucinate and it's annoying. A robot can hallucinate and it crashes into your bookshelf. That gap is what I find fascinating: the moment a perfectly reasonable plan from a perfectly reasonable model touches the real world and discovers friction, latency, a slipping gripper, a cable it didn't know was there.
I read the embodied-AI literature the way other people read thrillers. Manipulation policies, sim-to-real transfer, language- conditioned control, the new wave of vision-language-action models. I have a long mental list of things I want to build, and not enough weekends.
The honest version: I don't have a robot yet. I'm waiting for the right cheap arm to land, and the right weekend, and the right moment where I can spend two months making it pick up a pen and put it back down. When I do, expect bad video and better notes.