I talk to a personal agent from my phone and my Wear OS watch. I asked it for a game for my kid, then kept asking, and ended the evening with five games on my wrist. They are small and disposable, and I wouldn't have built them any other way.
A small tool that refuses any change which grows your code past its budget, and makes every raise an explicit, logged grant that shows up in the diff. How it works, how I run it on its own repo, and where it does not help.
A small Wear OS app on the Pixel Watch sends my voice to my personal assistant and reads the reply back. Most of the work was in one late session with coding agents, and most of the bugs were in speech recognition settings and sideloading.
Jarvis went looking for an open-source memory library to replace its own. Its first pick passed the test and turned out to be abandoned. Here is how it ended up on Graphiti, what the public benchmarks do and don't say, and why the real answer will come from replaying my own data.
My personal agent can now answer with a small web page, drawn in a sandboxed frame inside the chat bubble. Getting a game to show up was easy. Keeping it safe, on-theme, and running when the next message arrived was the actual work.
A one-way ratchet on code size stopped my agent-written codebase growing. Clamping it exactly onto the tree turned out to be the mistake, and the fix is three hundred lines of deliberate slack.
An orchestrator that talks to a human and dispatches long-running workers has an IO problem that is harder than its scheduling problem. Every design mistake I can point at turned out to be an IO mistake.
Two things make agents ship quickly — a work board a human can actually verify, and a gate cheap enough that everybody runs it. Both turned out to be the same problem, and it is a design problem rather than a discipline problem.
A taxonomy of every self-edit an autonomous agent made over seven weeks, with a pie chart, and the one category that ever made the codebase smaller.
Two replies to my agent carried the same facts. One got a thumbs up, one got a thumbs down, and the difference wasn't accuracy — it was whether the verdict survived contact with the reader.