testing-skills-with-subagents

You apply TDD to AI agent skills by running baseline scenarios without the skill to identify failures, writing the skill to make tests pass, and refactoring documentation to close loopholes. This RED-GREEN-REFACTOR cycle ensures robust agent behavior and compliance.

Pressure testing for AI agent compliance involves designing realistic scenarios with multiple combined pressures like tight deadlines, sunk costs, and authority figures. This process forces agent decision-making and captures rationalizations to iteratively refine skill documentation.

AI agents ignore skill documentation under pressure when loopholes exist in the rules. By capturing and addressing agent rationalizations for violating rules through simulated scenarios, you can iteratively refine the skill to ensure consistent enforcement.

To test if a skill enforces coding standards against corner-cutting, simulate scenarios where an agent might be tempted to violate rules, such as facing tight deadlines. Run baseline tests without the skill first to observe failures, then apply the skill to verify compliance.

No dependencies are required to run subagent skill testing. The process relies on using subagents to simulate scenarios with explicit options, real constraints, and direct calls to action, forcing agent decision-making to validate skill compliance.

You should avoid using TDD for skill validation when the skill does not enforce strict rules or require resistance to manipulation. If the skill performs simple data transformation without compliance risks, the overhead of pressure testing and rationalization capture is unnecessary.