Simular's Agent S surpasses human baseline on OSWorld benchmark

Simular announced that its open agentic framework, Agent S, has achieved a 72.6% success rate on the OSWorld benchmark for evaluating multimodal agents performing real computer tasks. This score exceeds the benchmark’s human-level performance of 72.36%, marking the first time an AI agent has surpassed the human threshold in this domain.

• Agent S reached 72.6% on OSWorld, beating the human baseline of 72.36%.
• The breakthrough was enabled by Behavior Best-of-N (bBoN), a method using multiple agents to select the best outcome.
• This represents significant progress from scores hovering around 20% just one year ago.
• Simular recently secured $21.5 million in funding and is piloting Microsoft’s Windows 365 for Agents.

Simular aims to make this technology widely accessible to unlock real use cases for consumers on real computers.