LegoEnv builds 5,465 agent environments from reusable atoms

A new arxiv preprint from Zhuosong Jiang and colleagues proposes building LLM agent training environments the way Lego sets are built: from reusable atoms. Their system, LegoEnv, assembles 5,465 verified environments across 223 applications by composing four pieces — intent, rule, data and server — and then evolves those atoms when an environment fails to give reliable feedback.

The authors frame the gap bluntly: "LLM agents learn by interacting with environments, yet existing environments are designed for specific tasks and fail to generalize to new ones." Prior methods, they argue, "generate each environment from scratch, neither reusing parts shared across environments nor carrying a fix from one environment to the rest." LegoEnv's pitch is that one round of evolution "raises the share of solvable environments from 71.6% to 88.3%," and that failures end up "concentrated in a few shared root causes" that can be patched once and propagated.

The downstream training numbers are mixed. On AutomationBench, supervised fine-tuning on LegoEnv data lifts Qwen3.6-27B from 16.0% to 32.0% — a doubling off a low base. Reinforcement learning with LegoEnv's own verifiers lifts Qwen3.8-27B from 37.6% to 40.3%, a 2.7-point gain. The paper also reports "steady gains with corpus size" and "transfer to unseen applications," though it does not publish per-application breakdowns in the abstract.

It was already circulating among a couple of the agent-infra researchers on our Who's Who radar within a day of posting, which tracks with a broader shift: the bottleneck in open agent training is increasingly the environment, not the model.