Programmatic skills may cut the cost of long-horizon AI agents

A new arXiv paper studies how agents can adapt to new domains while spending fewer model calls.

The authors argue that programmatic skills are especially cost-effective because they turn repeated sequences of actions into deterministic procedures. Instead of rediscovering the same behavior through trial and error, an agent can execute a learned program when a similar task appears later.

To test the idea, the paper introduces SpeedRunner, a coding agent that analyzes past trajectories and refactors skills for future tasks. The central hypothesis is that prior attempts contain enough signal to improve future performance even without replay or validation, if the agent can identify reusable structure.

The work focuses on cost reduction rather than only benchmark scores. That distinction matters for real deployments, where long-horizon agents can become expensive or unstable if every step requires fresh reasoning. The result suggests that durable, inspectable skills may be an important layer between prompting and full autonomy.