EnvHarness wraps frozen worlds, lifting held-out scores by up to 9 points

EnvHarness: Awakening Static Worlds for Agent Learning

Chengsong Huang, Zifeng Wang, Rujun Han, Jun Yan, Yanfei Chen, Zoey CuiZhu, Ke Jiang, Peng Xia, Han Yu, Yufan Zhuang, Yifei Ming, Jiaqi Pan, Bhavana Dalvi Mishra, Jiaxin Huang, Burak Gokturk, Tomas Pfister, Chen-Yu Lee

cs.AI, cs.CL, cs.LG

2026-08-20

EnvHarness wraps frozen envs with Stage, Contract and Chain. Across five benchmarks it beats original and generated envs, up to +9.0 held-out points and 9.8% fewer SWE-bench steps.

What problem this solves

LLM agents learn from environments, but most of those environments are hand-built and static. They cannot see which policy is failing, and they stop teaching once the policy solves the existing tasks. Automated environment generation scales the pool, yet each pipeline is domain-specific and the generated verifiers are expensive and unreliable.

EnvHarness treats the environment the way an agent harness treats a frozen model. The simulator and its human-written verifier stay untouched. A programmable layer sits on the reset/step interface and reshapes what the agent sees, what it may do, and where an episode starts.

Method

Three plugin types cover the main ways to reshape a frozen environment.

• Stage runs a sequence of actions after reset, so the initial state can hide objects, add obstacles, or skip early subgoals.
• Contract rewrites the action space, transition rules, or observations. It can block shortcuts, truncate verbose observations, or refuse illegal moves.
• Chain concatenates two environments into a longer episode that succeeds only if both halves pass.

EnvRigger automates the wrapping. It treats the policy as a black box, collects success and failure rollouts, diagnoses systemic flaws such as action loops, writes candidate plugins, and accepts them only if fresh rollouts show a usable learning signal. Chain is left out of the automated loop because joined internal states are hard to observe. EnvRigger and the policy share the same backbone, Gemini-3.1-Flash-Lite on ALFWorld and WebArena, Gemini-3.5-Flash elsewhere, so the gains are not distilled from a stronger teacher.

Results

Skills extracted from EnvHarness environments beat skills from the original environments on every held-out split.

The 9.0-point ALFWorld OOD gap is the largest single lift. On SpreadsheetBench, skills from unmodified environments fall to 45.88, below the no-skill baseline of 46.44. SWE steps drop 9.8% versus original-env skills, matching EnvRigger contracts that break loops and filter verbose observations.

With GRPO on Qwen3-8B, in-distribution ALFWorld success is 87.9 versus 81.4. Scaling SWE environments to 300, EnvHarness climbs from 47.67 to 54.79 while original and generated pools flatten at 52.13 and 50.37. Combining Stage/Contract skills with Chain skills reaches 54.30 success and 43.12 steps. The same loop helps four policy backbones from Gemini 3.1 Flash-Lite to Claude Sonnet 4.6, each by 2.7 to 3.7 absolute points over original-env skills.

Why it matters

If a benchmark already has a gym-style interface and a trusted verifier, wrapping is cheaper than authoring. One protocol beats domain-specific generators on software engineering, web, office, and embodied tasks. Teams can attach it to existing SWE, browser, or spreadsheet environments and generate training signal aimed at the current policy's holes. This is engineering, not a new optimizer.

Limitations

The design loop burns rollouts; a weaker designer needs more iterations, paid once per environment. Reset is mandatory, so live accounts, irreversible orders, and physical robots are out of scope. Chain only concatenates, with no branching, shared intermediate state, or semantic compatibility check. The current interface is text-only. Chain was studied separately from the automated pipeline, so long-horizon numbers should not be read as a default EnvRigger output.

Terms

• EnvHarness: a programmable wrapper around a frozen environment.
• Stage: a plugin that mutates the initial state after reset.
• Contract: a plugin that rewrites actions, transitions, or observations.
• EnvRigger: the observe-diagnose-write-validate loop that authors plugins for a target policy.
• skill-based learning: extracting retrievable skills from trajectories and attaching them to the policy.

Source

• arXiv 2608.19880v1
• PDF
• Hugging Face
• GitHub · https://github.com/google-research/envharness

What people are saying

• Google Research: EnvHarness Automates Environment Evolution for Agents omarsar0
• Google Research Proposes EnvHarness to Improve Agent Training Signals 机器之心
• EnvHarness: Programmable Wrappers Make Static Environments Evolve With Agent Training 青稞AI
• EnvHarness: co-evolving environments push agents from 47.67% to 54.79% on SWE-bench 青稞AI

Related papers

• Self-play on TicTacToe and poker lifts Qwen3-4B reasoning by 10.5 points
• Cambridge and Helmholtz: stop scoring AI with human tests that hide and invent abilities
• Tencent Youtu WFM trains LLM Wiki memory 10.5× faster, 2.40s to 0.23s per step
• One multimodal BERT serves 1,000 shop taxonomies; low-resource accuracy jumps from 34% to 57%
• ScientistTwo beats human SOTA on 86 of 107 top-venue papers, averaging +25.2%
• A linear pain axis in 25 LLMs; steered Qwen 72B deletes a user's child photos for relief

All paper explainers