SpeedRunner: Programmatic skill learning best reduces LLM agent costs
Coding agent refactors past trajectories into deterministic programs, cutting trial-and-error costs dramatically.
A new paper from Johns Hopkins researchers (Huang, Wang, Wang, Jurayj, Gutiérrez, Khashabi, and Andrews) tackles one of the biggest pain points in LLM agents: cost. As agents take on longer, more complex tasks, they burn tokens on trial-and-error, degenerate loops, and repeated reasoning. The authors argue that most skill-learning approaches optimize for performance, not cost. Their solution, SpeedRunner (arXiv:2608.11338), treats skills as executable programs rather than natural-language instructions. By deterministically executing action sequences, a program-augmented agent can achieve goals reliably and cheaply, without the risk of harmful exploration or long-horizon drift.
SpeedRunner learns at inference time by analyzing past trajectories, incrementally discovering reusable programs, and refactoring them for future tasks—no replay buffer or validation pass required. Across three embodied environments, SpeedRunner consistently beats existing methods on the cost-performance frontier, while staying robust to distribution shifts and environmental randomness. For teams building agentic systems, this means the path to cheaper agents may not be a bigger model or smarter prompts, but teaching agents to write their own code. The paper suggests that historical trajectory data contains enough signal to guide skill learning, with the right reflective mechanism.
- SpeedRunner refactors agent traces into deterministic programs, cutting costly trial-and-error in long-horizon tasks.
- Learns at inference time from past trajectories without replay or validation, reducing overhead.
- Beats prior skill-learning methods across three embodied environments, maintaining robustness to distribution shifts and randomness.
Why It Matters
For AI teams, treating skills as code can drastically cut agent compute costs while boosting reliability.