Intel / article

Better, Faster, Stronger: Programmatic Skill Learning Best Reduces Agent Cost

Source: arxiv.org
Author: Zixi Huang, Xiheng Wang, Andrew Wang, William Jurayj, Bernal Jim\'enez Guti\'errez, Daniel Khashabi, Nicholas Andrews
Date: Published Aug 13, 2026

Context

SpeedRunner explores whether agents can make repeated tasks cheaper by turning experience into reusable programs. It examines records of earlier action sequences and revises code-based skills for later tasks. The proposed advantage is that a program can specify a sequence directly, reducing repeated model-guided trial and error. The researchers evaluate this approach in three embodied environments.

The learning occurs in the 's external collection during use. The authors report that it achieves the frontier in learning and cost reduction across those environments while remaining robust to distribution shifts and environmental randomness. Executing a programmed sequence consistently does not mean the surrounding environment will always produce the same outcome.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.

Key quotes

“We argue that among all the different skill learning methods, those that view skills as programs can achieve the best cost reduction.”
Zixi Huang et al.

“By executing sequences of actions deterministically, a program-augmented agent can reliably and cheaply achieve goals that would otherwise require trial and error and risk degenerate behavior over long horizons.”
Zixi Huang et al.

“We hypothesize that past trajectories contain enough signal to guide skill learning, even without replay or validation, provided the agent can learn to analyze them.”
Zixi Huang et al.

“Across three different embodied environments, we show that SpeedRunner consistently achieves the frontier in learning and cost reduction while remaining robust against distribution shifts and environmental randomness.”
Zixi Huang et al.

Read the source arxiv.org