In October 2026 I shared a masterclass workflow for building, testing, and governing AI coding agents. I start by finding a working example in the open‑source Awesome LLM Apps repo, ask a coding agent to explain and run it, then experiment with live voice and camera input. I redesign the UI via a markdown file, use the “interview me” skill to clarify the idea, turn that dialogue into a formal specification, and finally build and test against that spec.
One model scored 23% in one harness and 52% in another on SWE-bench Pro.
Multi-harness RL trains the model inside the real harnesses, all at once, without touching their code.
Advent of Agents 3! 31 free days of building secure and governed AI agents with Google Cloud.
- Pre‑production checklist: edge cases, evaluations, observability, governance, cost.
- Decision: use multi‑harness RL because single‑harness scores are inconsistent.
- Decision: adopt Google Cloud governance framework for its structured, security‑first lifecycle.