Tmax: A simple recipe for terminal agents

2026-06-22

AI Summary

Current terminal-based agents suffer from the absence of efficient reinforcement learning training methods, high-quality interaction data, and reproducible baselines. This work proposes Tmax, a concise and effective training framework that constructs a large-scale, high-quality dataset of terminal interactions through a data generation strategy integrating difficulty control, role assignment, and validator diversification. Tmax employs outcome-reward-only reinforcement learning to train open-weight models without relying on intermediate supervisory signals. Experimental results demonstrate that the 9B-parameter model trained with Tmax achieves a 27% success rate on Terminal-Bench 2.0, surpassing prior models of substantially larger scale. Furthermore, the publicly released dataset exceeds existing counterparts in size by more than 2.5×, offering a valuable resource for future research in this domain.

Abstract

Terminal-using agents have quickly become the most popular downstream application of language models (LMs). Despite their prevalence, relatively little academic work has examined RL-based training of these models, likely due to difficult benchmarks, a lack of data, and a lack of simple baseline recipes. We present Tmax, the strongest open RL recipe for terminal agents to date, bringing open data recipes closer to the frontier. While simple, our recipe achieves 27\% on Terminal-Bench 2.0 with only 9B parameters, outperforming much larger models from prior work. Concretely, we generate data using a novel taxonomy, combining difficulty control, personas, and verifier diversification, which allows us to cheaply generate large amounts of terminal environments for RL and SFT training. We open-source our terminal dataset, which is over 2.5x larger than previously released terminal-agent datasets. We then train open-weight models using RL with our data, using a simple, outcome-only recipe. We release our data, models, and code as a strong baseline for future open academic work on terminal agents at https://github.com/hamishivi/tmax.

Problem

• terminal agents
• reinforcement learning
• language models
• benchmarking
• open data

Innovation

• terminal agents
• reinforcement learning
• data generation taxonomy
• difficulty control
• open-weight models