World-model agent replays environments for other agents

World-model agent replays environments for other agents — Trace2Env code released

Researchers Quanyu Long and colleagues published the paper “From Traces to Agentic Worlds” (arXiv:2610.06100) on October 5, 2026, and released the Trace2Env framework under Apache-2.0. Without training, it builds an immutable “world book” from recorded logs and uses a world-model agent to replay environments for other agents.

Compiled by Sergey KostenchukPublished 2026-10-10Updated 2026-10-10

2026-10-10 Coding HuggingFace

🛠 World-model agent replays environments — Trace2Env code released

Quanyu Long and colleagues published the paper “From Traces to Agentic Worlds” (arXiv:2610.06100) on October 5, 2026, and released the Trace2Env code under Apache-2.0. Without training, an immutable “world book” is created from dozens of recorded logs, and a world-model agent replays a live environment based on it for a task agent. On AgentWorldBench, the method achieves 75.94 on GPT-5.6-Sol compared to 69.51 for Direct Prompting.

🌍 Sandboxes for LLM agents can be built without a real environment: just logs from an unavailable, private, or legacy system are enough.

👤 The demo is open on Hugging Face Spaces: the terminal is fully replayed by an agent; a read file “disappears” after rm. The CLI demo works without a model and API key, and WALKTHROUGH.md helps build your own “world book” from logs.

Source 1: https://arxiv.org/abs/2610.06100 Source 2: https://quanyulong.net/trace2env/ Source 3: https://github.com/ruyue0001/trace2env

Sources