Google Releases EnvHarness to Make AI Agent Training Environments Adapt

Google Research has released EnvHarness, an open-source framework designed to make static AI-agent training environments adapt to the agent using them. Rather than replace a coding sandbox, website, or other simulator, it wraps the existing environment and changes the starting state, available actions, observations, or task sequence—while leaving the original simulator and success checker in place.

That distinction separates EnvHarness from approaches that generate entirely new training environments. The researchers’ premise is that creating a reliable environment and verifier is expensive, while a fixed task pool becomes less useful as an agent learns its patterns. EnvHarness instead aims to draw more training value from a smaller set of environments whose grading logic is already trusted.

Change the lesson, not the grader

The framework has three building blocks. Stage changes where an episode begins; Contract changes the interaction by filtering actions or altering what the agent sees; and Chain connects tasks into a longer episode. A coding-oriented Contract, for example, can stop an agent from submitting a patch before it runs tests. The source repository and its human-written tests remain unchanged, so the original tests still decide whether the patch is correct.

Choosing those modifications is the job of EnvRigger, a companion system that follows an observe, diagnose, write, validate loop. It examines repeated agent rollouts for failure patterns, creates candidate changes, then tests whether the altered task remains useful and solvable. The research describes validation as a safeguard against modifications that make a task impossible or trivial—not proof that every generated wrapper will improve a production agent.

A benchmark gain, with a practical boundary

In experiments across five benchmarks, agents trained with EnvHarness environments outperformed agents trained on unchanged environments, with gains of up to nine percentage points on held-out tasks. On SWE-bench Verified, the researchers reported that expanding the EnvHarness pool to 300 environments lifted the base agent from 47.67% to 54.79%; training on the same number of original environments reached 52.13%.

The same SWE-bench experiment also shortened average trajectories from 55.01 to 49.61 steps. Those reported results suggest that targeted constraints can teach an agent to avoid some inefficient habits. But they come from the researchers’ benchmark setup, in which trajectories were turned into reusable skills; EnvHarness itself does not train the agent. A deployment still needs a separate learning mechanism, such as skill extraction, fine-tuning, or reinforcement learning.

The environment must be safe to replay

EnvHarness is most suited to resettable digital sandboxes, such as coding environments, web-automation test systems, and tool-use simulations. Teams must connect an environment through a Bridge interface, and EnvRigger adds compute work because it repeatedly runs agents to diagnose and validate changes. The researchers caution against using the diagnostic loop directly on live production databases, real customer accounts, or physical robots, where resets may be costly or side effects cannot be undone.