From Signal to PR: Anatomy of a Self-Improving Agent — Jason Lopatecki, Arize

Arize’s proposed loop starts from an event or periodic trigger, gathers traces, logs, eval results, and repository context through skills, then opens an issue or PR. The key input is evidence about the actual production code path, not a blind request to inspect the codebase.

Instrument agent systems so a coding agent can retrieve compact, relevant evidence inside its sandbox. Compose observability skills with repository access, choose the harness and sandbox explicitly, and make the reviewable artifact the output.

This is currently closer to a **responder-to-reviewer** shift than autonomous repair. Small fixes are the best fit; larger changes still need a human to steer them, and production access may require a **VPC-hosted sandbox** with carefully scoped connections.