Background Coding Agents: Predictable Results Through Strong Feedback Loops (Honk, Part 3)

Introduced background coding agents ('Honk') for automated software maintenance. Designed a verification loop with independent verifiers (e.g., Maven, build, test) that abstract away complexity and provide incremental feedback to the agent. Implemented an LLM judge to veto functionally incorrect or out-of-scope PRs, vetoing ~25% of agent sessions and leading to course correction in ~50% of those cases. The agent operates in a sandboxed environment with limited permissions for predictability and security.