Tsinghua and Qwen team's Terminal-Universe turns agent trajectories into scalable terminal environments
rohanpaul_ai · x · 2026-09-10
A new Tsinghua + Qwen team paper flips the usual agent-data recipe: instead of imitating original runs, reconstruct the actual code workspace from a trajectory's tool-execution history and re-solve it.
- A trajectory is a single frozen demo; environments can be re-queried into many verifiable tasks with execution feedback.
- Terminal-Universe replays recorded file operations to restore pre-modification files, then a completion agent fills in missing files and dependencies.
- On the recovered workspace it both reconstructs the original task and synthesizes new ones, scaling along breadth (dependency mining) and depth.
Result: training on re-solved reconstructed workspaces yields better coding agents than imitating the original trajectories, easing the scarcity of executable environments for agent post-training.
More from coding & agent
- Armin Ronacher: GPT 6 Astra is impressive, but the 35-hour unattended software factory doesn't work — rseroter · 2026-09-10
- macOS tip: store secrets and tokens in Apple Keychain instead of env files — HankYeomans · 2026-09-10
- Dev ships six agents inside a self-hosted CRM instead of a chat window, shares lessons — PleasantBid4665 · 2026-09-10
- The Software Factory Stack: Open, Composable, Defined-in-Code — vikvang1 · 2026-09-10
- Browserbase: browser agents that write code beat pixel-clicking CUA models — adnan_hashmi · 2026-09-10
- Codex CLI 0.154.0 ships GPT-6-Astra, experimental worktree support — github-actions[bot] · 2026-09-10