Ramp’s first self-maintenance approach was a nightly agent running QA passes — sanity-testing core features, stress-testing recent PRs, probing for latent bugs. The limitation: the agent always followed the same paths. It was good at high-radius issues (things that affect many paths) but consistently missed narrow, situational bugs. This path dependency is an emergent property of unfocused agents — given a broad mandate, they converge on the same exploration strategy each run.
This explains why Autonomous coding loops need small stories and fast feedback to work works at a mechanistic level: small, specific stories force agents onto novel paths that a broad mandate would never reach. A focused mission (“check if the new currency formatter handles negative values”) explores a path that “run QA on the whole system” never would. Ramp’s solution was architectural: instead of one unfocused agent doing broad sweeps, they triggered targeted agents from specific monitors with specific alert context. The fix for path dependency is not a smarter agent but a more specific mission — which connects to Declarative beats imperative when working with agents: the constraint (specific success criteria) is what makes the agent effective, not the freedom.