testing-skills-with-subagents

Test-Driven Development (TDD) for skills uses a RED-GREEN-REFACTOR cycle: run scenarios without the skill to document agent failures, write the skill to address those failures, then refine it to counter new rationalizations. This approach exposes loopholes agents might exploit and makes skills bulletproof.

Use subagent testing to run pressure scenarios and capture the exact rationalizations agents use to bypass rules. Document each failure, craft a skill to prevent it, then re-test with variations. Iterate until the agent complies even under high-stakes pressure.

Establish a baseline by running your scenario without the skill and logging failures. Apply the skill and verify compliance. Then stress-test with pressure cases and refined rationalizations. Use measurable pass/fail criteria to confirm the skill blocks all documented workarounds.

Yes. TDD-style subagent testing applies RED-GREEN-REFACTOR cycles to skill development, ensuring agents follow rules consistently in high-stakes situations. It guarantees compliance by systematically identifying and eliminating behavioral loopholes before deployment.

You need a baseline scenario, scripted pressure cases designed to trigger agent rationalizations, rationalization logging to capture exact workarounds, and measurable pass/fail criteria. Start with the RED phase—run without the skill to document what agents do wrong.