Google released EnvHarness, a layer that trains agents against their weaknesses

An agent that keeps passing a benchmark eventually stops learning from it, because the start states, the rules and the feedback never change. Google Research has released EnvHarness, a wrapper that sits between the agent and the environment and lets a second loop rewrite the environment in response to the agent's own mistakes, without touching the agent or the evaluator.

The code is at google-research/envharness under Apache 2.0, with the paper on arXiv and a project page at envharness.com that lets a reader try the layer in the browser. InfoQ reports that Google's researchers describe their goal as making it practical for engineers to train an agent in an environment that keeps adapting, rather than one that is written once and left alone.

The design is three plug-in layers. A Stage sets up the task, a Contract defines the rule the agent must follow, and a Chain joins steps into one. All three sit on top of the existing environment's reset-and-step interface and leave its built-in verifier in place, so the original tasks and their pass or fail checks still work. Writing a new wrapper does not require writing a new benchmark.

A second component, the EnvRigger, is what makes the layer adaptive. It is an LLM loop that reads the agent's own session logs, finds the mistakes the agent keeps making, writes new wrapper layers that target those mistakes, and tests them against fresh runs. Running the loop over and over produces what Google calls a co-evolving environment: the agent gets better at the task, and the task keeps getting harder to find the next thing the agent cannot do.

Google says the paper measured the layer on WebArena, SWE-bench Verified and ALFWorld with four models: Gemini 3.1 Flash-Lite, Qwen3.6 27B, Gemini 3.5 Flash and Claude Sonnet 4.6. Every one of the four gained between 2.9 and 3.7 percentage points in task score with the EnvHarness layer added. The agent's own weights and prompts were untouched: the gain came from the environment adapting to the agent.

The repository has 645 stars under Apache 2.0 and the project page lets a reader add a wrapper layer and see the agent run against it in the browser, which saves setting up a local training environment to try the idea.

A team already running evals against a frozen harness has one honest question to answer first: is the agent still learning, or has its score stopped changing because the environment has nothing new to say. If the harness has not changed for a while and the agent's score has not either, the environment is probably what is holding the score steady, and this is the one piece that can raise what the harness asks of the agent without rewriting the harness or the evaluator.

Source

InfoQ, by Sergio De Simone. Primary source: envharness.com, google-research/envharness, arXiv paper.