Agents as a Service
/ blog
Experiments, findings, and lessons from building software with agents.
Coding agents remember little from one session to the next. The code tells them what a system does, but not why, and not which ideas were already tried and dropped. A decision log fills that gap: one file per decision, an index, a short section in the agent's instructions, and a check. How it worked while agents built a 3D space game, where it helped, where it hurt, and the instructions to copy into your own project.
For the second time, we gave an AI agent its own source code and told it to leave behind a better version of itself. Last time the only thing we learned for certain was that our referee could not tell an improvement from a copy. So we rebuilt the referee, set four lines of the agent loose against each other for 150 rounds, and this time scored each version on the children it produced. What they built with that freedom, what they claimed about it, and what a program does when it is finally selected for getting better at getting better.
We gave an AI agent its own source code, two days, and one instruction: leave behind a better version of yourself. Two hundred and twenty-one rewrites later it could search the web, read documents, query databases, and patch itself in transactions, all judged by its own earlier selves. This is the story of what it built, what its judges were actually looking at, and how we would run the experiment again to get closer to a real singularity.
