Computer-Use & Browser Environments

GUI agents fail where the desktop gets real

Computer-use models demo well on toy pages and stall on real work: dense enterprise interfaces, state that persists across applications, workflows that span a browser, a spreadsheet, and a desktop client in one episode. Our computer-use environments are built from exactly that work, elicited from the practitioners who do it daily and authored on the real software.

The operational problem is the product problem

Teams evaluating desktop agents at scale describe the same bottleneck: after every run the environment gets polluted, and restoring a clean state across thousands of daily jobs is its own engineering program. Our environments treat that as a first-class requirement. Deterministic seeds, per-episode resets, and snapshot restore come standard, so the eval fleet runs without human intervention.

Multi-user scenarios, because work is collaborative

Single-user benchmarks miss how office work actually happens. Episodes can simulate the other people in the loop: a manager who needs an approval, a candidate who stops responding, a colleague whose edits conflict with the agent’s. Collaborative capability is one of the least mature areas in agent evaluation across the industry, and it is where our long-horizon, multi-stakeholder task design goes deepest.

Verified like everything else we ship

Screen-level realism means nothing if the grading is soft. Every computer-use environment carries the same reward discipline as the rest of the catalog: deterministic verifiers grounded on the authoring expert’s own recorded run, state assertions against the application’s database, and model-graded criteria capped and labeled.