Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments

TL;DR

Reconstructs executable environments out of the agent trajectories that already exist, by replaying the file operations a trajectory recorded to restore each file to its pre-modification state, then having a completion agent supply the missing files and dependencies. The argument for doing so is the useful part: a trajectory is a single frozen demonstration, while an environment can be re-queried into many verifiable tasks and returns execution feedback (source).

Authors & Org

Not published in anything read. The snapshot carries title, abstract and upvote count but no author list, and arxiv.org answers EGRESS_BLOCKED from this run's sandbox. HuggingFace Daily Papers, 2026-09-05, 213 upvotes — the third-highest entry in that snapshot; arXiv publication date 2026-09-03 (source).

Method

The stated asymmetry: terminal-based code agents are prevalent and their trajectories have accumulated at scale, while realistic executable environments remain scarce — and environments are what agent post-training actually consumes (source).

The observation the method turns on: the tool-execution history inside a trajectory exposes the structure and contents of the environment it ran in, so the environment can be recovered from the trajectory rather than generated from scratch.

Reconstruction

1. Replay the file operations recorded in the trajectory to restore each file to its state before the agent modified it → a partial workspace
2. A completion agent supplies the missing files and dependencies
3. On the recovered workspace, both reconstruct the original intent task and synthesise entirely new ones

Scaling, on two axes

• Breadth — mine directional dependency relations between related environments and synthesise cross-workspace queries spanning multiple codebases, on the argument that developers routinely work that way
• Depth — extend a single-turn query into a multi-round session capturing iterative user feedback and requirement refinement, driven by a user agent

Results

The abstract carries no figures. No benchmark, no baseline, no environment count, no success rate, no model name. Everything above is method. This page records the mechanism and explicitly does not attribute any performance claim to the paper, because nothing read makes one.

That is unusual for a 213-upvote entry and is recorded rather than filled in.

Significance

It is the supply side of Post-Training Scaling. That page holds the claim — named as a scaling law by Jie Tang in August, with GLM-5.3 as the experiment — that the next increment of capability is bought after pre-training, from RL and generated training environments and the harness a model is trained inside. Environments have been the named ingredient with no published means of production. This is a proposed means of production, and its input is a byproduct every lab running coding agents already has in quantity.

It also sits directly against this week's harness cluster, from the other side. WHALE: A Simple Recipe for Joint Harness-Weight Optimization, HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness? and Aspire: Can Models Self-Evolve from Vague Goals? all optimise the loop the agent runs in; Terminal-Universe manufactures the world the loop runs against. Neither line cites the other, and together they describe an agent post-training stack whose three components — weights, harness, environment — are each now being scaled by somebody, independently.

The nearest existing page is EnvHarness: Awakening Static Worlds for Agent Learning (arXiv:2608.19880), and the distinction is worth keeping: reconstructing an environment from a recorded execution is a replay problem, not a generation problem, which is why the fidelity question below is answerable in principle.

Open Questions

• No numbers at all. Whether models post-trained on reconstructed environments outperform models trained on the trajectories they came from is the entire practical question, and the abstract does not touch it
• How faithful is a completed workspace? The replay restores files the trajectory touched; a completion agent invents the rest. A dependency the completion agent guesses wrong produces an environment that is executable and not the one the trajectory ran in — and every task synthesised on it inherits that
• Does this launder trajectory bias into environment bias? Environments are recovered only where trajectories exist, so the distribution of available environments is the distribution of tasks agents have already been asked to do. That is the opposite of the coverage argument the method is motivated by
• What licenses the recovered code? Trajectories from real codebases reconstruct real codebases. Nothing read addresses provenance
• Is the multi-round "user agent" evaluated, or assumed? The depth axis simulates the human whose feedback is the thing being modelled

Cite

Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments. arXiv:2609.04148, 2026-09-03. Recorded from HuggingFace Daily Papers, 2026-09-05 (source).