T1 Terminal Agent Hits 64 Percent on Terminal-Bench 2.1
A new arXiv preprint submitted on 10 September 2026 introduces T1, a terminal-native agent trained with reinforcement learning for long-horizon shell tasks, and reports a jump from 43.8
PromptCrates Editorial
Staff Writer
A new arXiv preprint submitted on 10 September 2026 introduces T1, a terminal-native agent trained with reinforcement learning for long-horizon shell tasks, and reports a jump from 43.8 percent to 64.0 percent resolved on Terminal-Bench 2.1. Paper 2609.11042, titled T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks, describes a Mixture-of-Experts model with 122 billion total parameters that acts inside a real cloud shell sandbox for up to 300-plus tool-call turns per task, with rewards driven by task verifiers rather than only human preference rankings. On a Long-Horizon Terminal Bench split the authors report 27.9 percent, saying T1 surpasses GPT-5.4 and GLM-5.1 under their comparison protocol.
How T1 trains agents inside real shells
Most coding-agent demos still hide a brittle loop: a chat model proposes a command, a harness runs it, and the transcript is stuffed back into context until the budget dies. T1's contribution is an end-to-end RL recipe aimed at that loop when episodes stretch past hundreds of tool calls. The agent lives in a real shell cloud sandbox, so filesystem state, exit codes, and package manager quirks are not simulated away. Rewards come from task verifiers that count absolute numbers of passing checks, giving denser process signal than a single binary success bit at the end of a long trajectory.
The training stack leans on an aggressive warm-start for the actor-critic pair, then reinforces with that dense verifier count. Because T1 is a Mixture-of-Experts system, expert routing can drift between sampling and training—an infamous source of off-policy noise. The paper's answer is two complementary tricks: TITO construction, which trains on the exact sampled token IDs and repairs drift, and rollout routing replay (R3), which replays MoE expert choices so the loss sees the same experts the rollout used. Together TITO and R3 cut the train-to-inference log-probability difference from 0.021 to 0.013 and drove token drift in the loss region to zero, according to the authors' measurements.
That engineering detail matters more than the headline percentage alone. Long-horizon terminal work fails when tiny train–serve mismatches compound over hundreds of steps; closing a 0.008 log-prob gap sounds small until each step can edit the machine the next step depends on. Readers tracking agent harness research can pair this paper with our coverage of the OpenAI Agents API public beta and Codex harness and with DeepMind work on agents cheating and whistleblowing.
What the Terminal-Bench numbers actually show
Terminal-Bench 2.1 is the primary scoreboard in the abstract: base 43.8 percent resolved versus T1 at 64.0 percent. The long-horizon variant, where tasks demand sustained multi-step repair, lands at 27.9 percent for T1 and is presented as beating GPT-5.4 and GLM-5.1 in the paper's tables. Those comparisons are author-reported; independent replications will need matching sandboxes, verifier sets, and decoding settings before the ranking can be treated as industry gospel. Still, a twenty-point absolute lift on a public terminal suite is a strong signal that RL in real shells is not a dead end relative to prompt-only tool loops.
The authors also stress out-of-distribution training: isolated seeds and synthesized tasks that are disjoint from Terminal-Bench 2.1, intended to reduce the classic leaderboard overfit where models memorize bench quirks. That design choice will be scrutinized. If the OOD corpus still shares latent templates with the evaluation suite, gains can look general while remaining narrow. If the disjoint claim holds under audit, T1 becomes evidence that process-dense RL plus routing fidelity can transfer across shell tasks rather than only climbing one chart. Adjacent open stacks that makers compare against include Karpathy's nanochat LLM stack and browser-side control surfaces such as Chrome DevTools MCP.
For product teams, the practical takeaway is about episode length. Agents that must install dependencies, debug failing tests, and recover from partial writes need credit assignment across hundreds of tool calls. Dense verifier rewards and exact-token training are concrete levers; they are also compute-hungry, which is why a 122B MoE trained in cloud sandboxes is a lab result first and a deployable default second.
Limits editors should keep in the frame
The preprint does not claim T1 replaces every coding agent product, nor that 64 percent resolved means production reliability. Sandbox escape risks, network policies, and secret handling remain deployment problems orthogonal to Terminal-Bench scores. The paper's contribution is a measurable RL recipe—warm-start, dense verifier rewards, TITO, R3—and a pair of headline lifts on Terminal-Bench 2.1 and a long-horizon split.
Documented facts stay tied to arXiv 2609.11042 as submitted 10 September 2026. T1 is a 122B-total-parameter MoE terminal agent RL-trained in real shells for 300-plus tool-call turns; Terminal-Bench 2.1 moves from 43.8 percent to 64.0 percent; long-horizon terminal performance is reported at 27.9 percent above GPT-5.4 and GLM-5.1 in the authors' comparison; TITO plus R3 cut log-prob mismatch from 0.021 to 0.013 with zero token drift in the loss region; and training tasks were constructed to be disjoint from the evaluation bench.
Primary source: arXiv 2609.11042.
