testing-skills-with-subagents

Test skills with subagents under pressure scenarios—combining time constraints, sunk-cost pressure, and authority pressure—to surface rationalizations. Document every loophole agents find, refactor the skill to close it, and re-test until the skill consistently enforces intended behavior without workarounds.

TDD for skills applies Red-Green-Refactor cycles to skill documentation: write a failing test scenario where agents rationalize noncompliance (RED), refactor the skill to block that rationalization (GREEN), and verify repeatable compliance (REFACTOR). This ensures skills withstand pressure and remain robust across deployments.

Combine 3 or more pressure types—time limits, sunk-cost framing, authority constraints—in a single scenario. Run the skill with subagents under these pressures, capture rationalization attempts in logs, and identify which skill language or constraints failed. Use these findings to tighten skill wording and enforcement.

Yes. Run baseline tests without the skill to establish noncompliant behavior, then test with the skill under realistic pressures. Verify GREEN outcomes (compliant behavior) and REFACTOR outcomes (documented skill improvements) before deployment to ensure consistent, compliant agent performance.

Log every rationalization agents attempt during pressure tests. For each loophole, update skill language to explicitly block that rationalization path, then re-test with subagents. Repeat until no new rationalizations emerge and compliance metrics stabilize across all pressure combinations.

No specific framework is required, but clear skill documentation improves testability. The process itself enforces compliance by iteratively tightening skill language based on actual agent behavior under pressure, making vague or exploitable instructions visible and fixable.