The Trust Ladder

Autonomy without governance degrades quickly. An agent that can write code, send messages, or edit production configuration is useful only if the boundaries are clear. We use a trust ladder with three levels. Level 1 covers reversible work with low blast radius: drafting, analysis, local code changes, test execution, and documentation. Level 2 covers actions that affect shared systems but are still recoverable: merging approved changes, publishing content, or updating routine automation. Level 3 covers irreversible or high-risk actions: payments, secrets, legal commitments, and production changes with meaningful customer impact.

The value of the ladder is not bureaucracy. It is consistency. Agents do better when approval rules are concrete instead of implied. A vague instruction like "use judgment" sounds flexible but usually creates hesitation in safe cases and overreach in dangerous ones. We want the opposite: fast movement inside clear constraints.

Once the ladder is explicit, review becomes targeted. We do not supervise everything. We supervise the transitions where the cost of error jumps, and we let the rest of the system run.