Terminal agents build their own RL environments: up to 25x more valid ones, 75% cheaper each

Reinforcement-learning environments for training terminal agents no longer have to be designed only by human researchers; agents generate them from the rollouts of the solver being trained. Such environments were previously mostly designed by hand, so their supply grew only as fast as researchers could work. Against prompt-only generation, AutoEnvScaling raises the valid-environment rate by up to 25 times and lowers cost per valid environment by up to 75%.

Environment generation is carried out as a terminal task, with a proposer agent building new environments and the solver's new rollouts guiding the next round. This data flywheel puts environment generation and task solving into one terminal loop instead of two separate processes. Training Qwen3.5-35B-A3B on AutoEnvScaling environments from GPT-5.6-sol improves its Terminal-Bench 2.1 score by 4.3 points over the matched Tmax baseline.

One model can act in both roles at once, as environment generator and as solver, so the two roles co-evolve. Earlier pipelines kept environment supply and solver training apart, with the model iterating in only one role. Qwen3.6-35B-A3B acting in both roles gains 12.2 points on Terminal-Bench 2.1 and raises its Terminal-Bench 4.0 score from 1.6 to 3.2.

This report covers the summary-level content of AutoEnvScaling: environment generation is treated as a terminal task, a proposer agent builds new environments from the solver's rollouts, and the pipeline generates environments for software engineering, computer use, professional work, and GPU kernel programming. It is aimed at researchers and engineers working on reinforcement-learning training and environment supply for terminal agents. The source states no next experimental step.

The source presents results in summary form and does not describe the implementation, verification, or scale of environment generation. The valid-environment rate and cost figures are upper bounds relative to prompt-only generation, and the Terminal-Bench scores come from specific model and baseline pairings; the source does not report behavior under other settings.