Harnessing LLM Agents with Skill Programs

HASP turns reusable LLM skills into executable Program Functions that can intervene when agents reach failure-prone states.

The framework can guide or correct agent behavior at inference time, during post-training, or through iterative skill evolution.

In benchmarks on web search, math, and coding, HASP outperformed both baseline agents and trained methods.

The approach makes agent skills actionable rather than advisory, improving reliability on long-horizon tasks.