A differential oracle: making agentic code prove its own correctness

In a shift from kitchen to coding, the author shares their journey in creating trustworthy agentic systems, focusing on the evaluation layer. This system allows coding agents to not only demonstrate their capabilities but also to independently verify their own correctness, which is crucial when they interact with real codebases. This innovation addresses the challenge of ensuring reliability in complex systems where manual verification isn't feasible. The implications are significant, promising more robust and self-monitoring software in the future.