Autoresearch: automating the optimization loop with agents

Most engineering improvement is the same loop: change one thing, measure, decide whether to keep it, go again. How fast that loop turns has always been set by the person turning it. An agent can now run it unattended, against any target with a number attached — test suite duration, bundle size, build times, model quality. Join us to explore how the idea applies to machine-learning problems — starting from Karpathy's autoresearch, with a hands-on example — and which parts of it carry over to any other target.