InAgent achieves 90.2% on OSWorld: the computer-use agent gap narrows for the Chinese stack

A Chinese CUA becomes the first to surpass 90% on OSWorld. The gap with US frontier labs is now a harness gap, not a model gap.

In plain terms. A Chinese computer-use agent called InAgent has posted 90.2% task-success on OSWorld—the standard benchmark for agents that click, type and navigate a desktop. Pandaily reports it as the first Computer-Use Agent (CUA) to cross 90%, ahead of published scores from OpenAI, Google and Anthropic.

What was delivered

According to Pandaily on August 3, 2026, InAgent achieves 90.2% task success on OSWorld, including 100% on system tasks. This marks a milestone for the "computer-use agent" field, which measures an agent’s ability to operate an OS like a human (mouse, keyboard, windows, applications).

Our take

The headline isn’t the number—OSWorld has a soft ceiling and a gaming halo when optimized for the benchmark—it’s the trajectory. The gap between frontier labs (OpenAI, Anthropic, Google) and Chinese players is closing on CUAs even before the consumer use case stabilizes. Pandaily frames this as “harness engineering as the new AI competition frontier.” It aligns with what we’ve tracked on the harness-ops thread for a month: product performance hinges less on the model than on the surrounding tooling loop.

Under the hood

On OSWorld, hitting 90% typically requires a stack: (a) UI perception model (screen understanding + OCR), (b) planner (task decomposition), (c) executor (atomic mouse/keyboard actions), (d) agent memory + backtracking to recover from errors. The 100% on system tasks suggests the CLI/scripts layer is solid; the delta to reach 90% globally comes from multi-window tasks with UI drift—the case where the screen changes between clicks and the agent must re-perceive.

So what

For teams building operational agents: (1) don’t benchmark your stack on published scores—reproduce, or wait for independent reproduction; (2) measure your own perception vs. execution gap—this is where costs explode; (3) rigorously separate perception, planning, and action roles in your architecture, or model gains will never reach the product.

To watch

Publication of InAgent’s model and/or harness—otherwise 90.2% remains a slide-deck number; independent replication (benchmarks must be openly reproducible or they’re meaningless); cost per task in consumed tokens—a 90% CUA that costs $10 per task isn’t industrializable; support for non-Windows OSes.