Microsoft Research trained AI agents on synthetic environments — realistic fake apps for the agent to practise in. The headline result is good: a small model nearly doubled its success rate, from 36.5% to 67.1%, landing within striking distance of a far larger and more expensive model.
The result worth your attention is a different one, buried in the ablations.
The Finding That Should Worry You
When they trained a model on a shallow environment — one that looked convincing but where the workflows did not really work end to end — the model's accuracy dropped below its own untrained baseline.
Not "less improvement." Not "a smaller gain." Worse than if it had never practised at all.
What Deep Actually Means
Their definition is strict, and it is not about features. A deep environment is one where every workflow it claims to support can actually be completed end to end through the real interface — permissions behave, errors fire, and a sent message genuinely reaches its recipient.
Note what that rules out. Depth is not a feature list. It is a claim you have to prove by walking through the thing. Most systems described as "comprehensive" or "production-realistic" have never been held to that bar.
A convincing surface with nothing behind it does not teach nothing. It teaches something wrong, confidently.
The Cheap Lever Backfires
There is a second trap. Once a model has learned what a given set of environments can teach, running more repetitions against the same set stops helping and starts hurting real-world transfer. What keeps working is adding genuinely different environments.
So the cheap lever — more reps, same material — actively damages the thing you care about. The expensive lever — genuinely new material — keeps paying.
Any team under time or budget pressure gets pulled toward the cheap lever by default. This is direct evidence that the instinct fails on exactly the metric that matters.
Hold This Up To Your Own Org
Every company runs shallow environments and calls it enablement:
- Role-play with a colleague who folds at the first objection, so the rep learns that objections are easy.
- A certification you pass by rewatching the video and re-taking the quiz.
- A sandbox with three clean demo accounts and none of the messy data that breaks things in production.
- An onboarding checklist that certifies attendance, not capability.
Each of these produces a number that goes up. That is the danger. The number rising is what stops anyone from asking whether the practice was real.
One Test
For anything you call training, ask: can someone fail it in a way that costs them something?
If the answer is no — if the exercise cannot be genuinely failed — then it is not a simulator. It is a video with a quiz attached, and the honest thing to do is stop counting it as preparation.
The uncomfortable version of this paper's finding: some of what you currently spend on training is not neutral. It is teaching people that the job is easier than it is.
Source: Yash Pandya, Akshay Nambi et al. (Microsoft Research), “Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale,” arXiv 2607.28074, July 2026.
