Across nearly 6,000 runs covering two office automation benchmarks and three model harness stacks, this study shows that average task-success gains conceal a substantial regression cost. A regression is a task solved without a skill but failed after the skill is added, while a residual failure fails in both conditions. The authors identify skill-description osmosis, grounding displacement, and verification displacement as major regression causes. Their analysis suggests that reliable agent skills depend more on grounding and verification than on procedural instructions alone.
No heat snapshots are available in the last 24 hours.