Passion Fruit AI passed Terminal-Bench 4 and OSWorld 2 tests

The development team behind Maracuya AI presented the results of testing its agent system in specialized environments Terminal-Bench 4 and OSWorld 2. These benchmarks assess AI agents’ ability to perform complex tasks in a real operating system, including working with the console, managing file pipelines, and automating business processes.

During testing, the agent framework demonstrated a high degree of autonomy in solving practical tasks. Unlike standard chatbots, the system being tested had to do more than generate text: it had to produce working artifacts, such as fixing errors in server scripts, processing data from email clients, and consolidating financial spreadsheets. The results demonstrated the model’s ability to plan across multiple steps and interact correctly with the command-line interface.

OSWorld 2 is one of the industry’s most demanding benchmarks, as it simulates a real user’s working environment. Passing tests like these confirms that agent systems are ready to automate routine system administration and office tasks. Using such tools makes it possible to objectively compare the performance of domestic developments with their global counterparts under conditions that closely resemble real-world deployment.

These tests highlight the shift from simple text models to full-fledged agents capable of managing computing resources. Further development of such systems is aimed at improving command execution reliability and minimizing errors when working with file systems and external APIs.