testing-skills-with-subagents

To prevent agents from bypassing instructions, apply test-driven documentation by running realistic pressure scenarios to observe failures, then write targeted counters to close those loopholes. This process hardens compliance-sensitive skills against rationalization.

The RED-GREEN-REFACTOR cycle for agent skills involves running baseline scenarios to record failures (RED), implementing minimal counters to address those failures (GREEN), and iteratively updating rules to close remaining loopholes (REFACTOR).

You test agent skills against operational pressures by creating multi-pressure scenarios involving time, sunk cost, or authority, capturing the agent's rationalizations verbatim, and using those records to drive precise documentation counters.

Use test-driven documentation when authoring or hardening compliance-sensitive or discipline-enforcing skills where agents may adapt, rationalize, or bypass instructions under operational pressure. It ensures instructions hold up under realistic stress.

You verify agent compliance by pressure-testing updated skills, requiring agents to cite specific documentation sections during execution, and performing meta-questions to identify and resolve any remaining ambiguities in the instructions.

Pressure-based skill testing relies on running realistic scenarios to capture rationalizations, but it may not anticipate every edge case. Meta-testing and iterative refactoring are required to continuously identify and close remaining ambiguities.