An Agent That Learns the Job
Real work is more detailed than anyone can describe upfront. How we treat an agent as a learning system: keep the traces, turn accepted work into evals, cluster rejected work into rules, and earn trust one reviewed change at a time.