Chinese AI Agent 'InAgent' Achieves Record 90.2% Success Rate on OSWorld Benchmark
InAgent, an AI agent developed in China, has achieved a groundbreaking success rate of 90.2% on the OSWorld benchmark for computer-use agents. This marks the first time an agent has surpassed the 90% success threshold in this domain. Notably, InAgent demonstrated 100% success on system-level tasks within the benchmark. The achievement positions InAgent ahead of record-setting agents from major AI research labs such as OpenAI, Google, and Anthropic. The OSWorld benchmark evaluates an agent's ability to perform complex tasks within a simulated computer environment. The rapid advancement in this area highlights 'harness engineering' as a new frontier in AI competition.
The successful development of InAgent and its surpassing of existing benchmarks set by major Western AI labs suggests a rapidly intensifying global competition in AI agent capabilities. This achievement underscores the growing significance of 'harness engineering'—the development of robust frameworks and evaluation systems for AI agents—as a critical determinant of progress. As AI agents become more adept at complex tasks, the focus will likely shift towards their real-world applicability, safety, and ethical deployment. Future developments may involve more sophisticated benchmarks that test not only task completion but also adaptability, resource efficiency, and human-AI collaboration, reflecting the evolving demands of the AI era.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.
