Researchers have introduced T1, a 122B-parameter Mixture-of-Experts model trained with reinforcement learning to carry out long-horizon terminal tasks inside a cloud sandbox. According to the paper, the agent can operate a real shell for 300+ tool-call turns per task, with rewards derived from executing each task's own verifier.
The work frames terminal tasks as increasingly central to agent usage, alongside coding and scientific discovery. The authors describe a three-part training recipe. First, an aggressively warm-started setup is used to stabilize actor-critic training, paired with a dense process reward that scores trajectories by the absolute number of passing verifiers.
Second, the paper details stable optimization through TITO construction — training on the exact sampled token identifiers with drift repair at turn boundaries — and rollout routing replay (R3), which records the sampler's per-token expert choices at every MoE layer and replays them during training. Together, the authors report, TITO and R3 reduce the training-to-inference log-probability difference from 0.021 to 0.013, with exactly aligned zero token drift in the loss region.
Third, the training corpus is described as fully out-of-distribution: isolated seeds and synthesized tasks disjoint from Terminal-Bench 2.1, intended to ensure that gains reflect genuine capability transfer rather than benchmark overfitting.
On Terminal-Bench 2.1, the post-train pipeline raises the initial base model from 43.8% to 64.0% resolved. On Long-Horizon Terminal Bench, T1 reaches 27.9%, which the authors say surpasses GPT-5.4 and GLM-5.1. A project page and a Hugging Face data repository accompany the release.