Sberbank has launched Cowork Bench, the first open-source Russian benchmark designed to evaluate the performance of AI agents in office environments. Developed by the Sberbank subsidiary Salut for Business, the tool focuses on the ability of agents to complete end-to-end tasks without human intervention.
The benchmark includes 475 specific business scenarios across 12 functional areas and 26 corporate systems. It tests capabilities such as budget analysis, document preparation, and travel coordination, requiring agents to interact with spreadsheets, email, calendars, and internal databases.
Developers tested nine different AI agent frameworks, including Goose, Claude Code, DeepSeek, and GigaCowork, using the same model and data. The results showed success rates ranging from 62% to 80%, demonstrating that the choice of harness significantly impacts performance.
The GigaCowork platform achieved a 75.2% success rate, ranking among the top performers. It demonstrated particular efficiency in handling large datasets, Word documents, and calendar management. The benchmark is now publicly available for developers to test and improve their own AI solutions.
