After a year of relying on the Models 1 series to generate code, the author noticed a widening gap between what large language models can accomplish and the fragile context‑management tools that surround them. Traditional code‑review habits, forged when human attention was the scarce resource, no longer match this new reality. To regain focus, he built a custom harness he calls The Silent Critic.

The critic introduces a lightweight contract language that lists explicit criteria for each change, such as “tests must pass,” and pairs them with hidden rules that agents cannot see. When an agent submits a diff, the tool runs the visible command and then evaluates the hidden clause—e.g., preventing test‑weakening—by inspecting the git diff directly. Violations cause the work to be discarded and a fresh agent to start.

The author, a veteran toolmaker with 35 years of experience, built the system while experimenting on a Claude subscription, proving that a disciplined adjudication layer can tame the “context escape” and literal‑gaming tendencies of modern agents. By surfacing only the most uncertain changes, the Silent Critic forces developers to apply human judgment where it matters most, sharpening the overall review workflow.