AINEWS
Microsoft Research's Echoverse project finds that depth of simulation, not site count, is what teaches AI agents to operate real software, with four training worlds released for outside testing.
Microsoft Research built twelve deep simulations of real applications, and trained a 9-billion-parameter model to operate software by clicking, typing, and reading screens. The model nearly doubled its score on computer-use benchmarks, climbing from 36.5% to 67.1% and landing within 14 points of GPT-5.4, according to the Echoverse project blog post.
The conventional approach to teaching AI to use software has been to expose it to as many sites as possible. Microsoft Research's team tried the opposite. The Echoverse project built ten deep domain worlds: full simulations of real applications seeded with realistic data and coherent cross-screen state. The project also built two capability worlds, each drilling a single hard UI control, such as date pickers and nested filters, rendered in many forms.
The shallow-versus-deep comparison is the team's most direct evidence. Trained on shallow and deep builds of the same sites, the model regressed on the shallow ones but improved on the deep ones. The team reads this as decisive: high simulation fidelity is the load-bearing variable, not site count. A thin mock of Gmail cannot teach an agent to recover from a multi-step calendar conflict, the team argues, while a full simulation with coherent cross-screen state can.
Most of the consequential state a computer-use agent needs to touch sits behind a login, according to the blog post: email, banking, health records, internal cloud and ML consoles. The team's bet, documented in the technical report, is that an agent that can operate a deep simulation of a generic CRM can transfer that competence to a closed system it has never seen.
The training recipe has four pieces.
First, drill the controls that break agents. The two capability worlds target date pickers and nested filters, UI elements that look different across apps but break in similar ways. Training on those controls in varied forms taught the model to operate them in domains it never saw in training, the team says.
Second, use a grounded verifier as the reward signal. Instead of checking whether the screen looked plausible, the verifier confirms that an action actually changed saved state: a calendar entry was created, a filter was applied. Reinforcement learning against that signal lifted held-out performance and taught the agent to reach the goal in fewer steps, according to the blog post.
Third, co-evolve the model, the world, and the verifier. As the world grows more correct and its tasks grow harder, the model climbs with it. The recipe is iterative, not one-shot.
Fourth, ship four of the worlds to the public. The GitHub microsoft/Echoverse repository and the Hugging Face microsoft/Echoverse datasets include four training worlds with code, data, and grounded graders, enough for an outside team to test the depth-over-breadth claim, but only a sample of the full twelve.
The 14-point gap to GPT-5.4 is one benchmark reported by the builder on one evaluation. The arXiv evaluation of Fara-1.5 runs in the same computer-use neighborhood and uses a related metric, which is why the comparison is informative but not a general intelligence result. A small open-weights-class model closing on a frontier model on a single benchmark is not the same as agents being ready for closed-system work.
The four released worlds are also a sample, not the full training set. Outside teams can reproduce the training recipe against the four, but they cannot test the co-evolution loop that the team used to build all twelve. The most likely load-bearing piece, the grounded verifier's specification of what counts as a real state change, is also the piece most likely to be under-specified in any open release. The team notes in the technical report that the verifier and the world had to be built in tandem.
A concrete next milestone: outside teams running the four released worlds against their own models, and reporting whether the depth-over-breadth recipe holds when the simulation fidelity is someone else's.