Terminal-Universe recycles 37,300 agent workspaces

Terminal-Universe reconstructed 37,300 task-sufficient environments from existing terminal-agent trajectories; fine-tuning Qwen3.5-27B on them improved Terminal-Bench 2.1 by 11.9 points. The paper’s results treat execution histories as reusable workspaces rather than frozen demonstrations, and report a 13.8-point gain on multi-round EvoCode-Bench. Teams training coding agents should compare this against the agent-skill retrieval bottleneck: trajectory archives can become training environments, but only after task and dependency reconstruction is verified.