Every agent framework in production today uses declared authority. The developer writes a config. The config is evaluated at deploy. The agent runs with whatever was declared, until the developer remembers to narrow it. It is the model that fails as the number of agents and the variety of their tasks grows.

This post introduces the conduct engine pattern: batting averages over decisions, regression detection, automatic de-escalation, and why the math of a conduct engine is actually tractable when the underlying governance layer captures decision frames for every action.

Publishing April 29, 2026. All posts.