T1 agent improves long task solving with 122B parameter model
T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks
Machine LearningArtificial Intelligence
Summary
Long tasks like coding or scientific discovery are hard for AI because they require many steps. The authors developed an AI model called T1 that works in a real computer shell, running many commands in sequence to complete tasks. They trained T1 using special techniques to make learning stable and tested it on tasks not seen before to prove it can adapt. T1 worked better than previous AI models on these long and complex tasks, showing progress in teaching AI to manage multi-step problems.
What this means in practice
- •For software developers: Automate complex coding tasks by using an AI agent that interacts with development tools in realistic environments to complete multi-step programming assignments.
- •For cloud platform engineers: Integrate reinforcement-learned agents like T1 to manage and automate extended command sequences in cloud-based sandbox environments.
Authors
Junyao Yang, Yucheng Shi, Zhongzhi Li, Ruhan Wang, Zongxia Li, Haitao Mi, Leowei Liang
Abstract
Agent usage is shifting toward long-horizon tasks such as coding and scientific discovery, among which terminal tasks are especially important. We introduce T1, a Mixture-of-Experts model of 122B total trained with reinforcement learning, operating a real shell in a cloud sandbox for up to 300+ tool-call turns per task, rewarded by executing each task's own verifier. We provide a comprehensive recipe: First, an aggressively warm-started to stabilize actor-critic training, with a dense process reward scoring trajectories by the absolute number of passing verifiers. Second, stable optimization through TITO construction, training on the exact sampled token identifiers with drift repair at turn boundaries, and rollout routing replay, recording the sampler's per-token expert choices at every MoE layer and replaying them during training. Third, fully out-of-distribution training corpus: isolated seeds and synthesized tasks disjoint from Terminal-Bench 2.1 ensures gains reflect genuine capability transfer over benchmark overfitting. Together, TITO and R3 cut the training-to-inference log-probability difference from 0.021 to 0.013, with exactly aligned zero token drift in the loss region. On Terminal-Bench 2.1, our post-train pipeline raises initial base model from 43.8% to T1 with 64.0% resolved. On Long-Horizon Terminal Bench, T1 reaches 27.9% and surpasses GPT-5.4 and GLM-5.1.