Workplace Assistant Agent Benchmark with 690 Tasks

A 2026 benchmark from NVIDIA provides a sandbox environment for evaluating multi-step agentic task execution. It contains 26 tools and 690 tasks simulating common business activities like sending emails and scheduling meetings. The dataset is part of the NVIDIA NeMo Gym framework for building reinforcement learning environments.

Use Cases

• Benchmarking agent performance on the 690 provided workplace tasks using the 26 available tools.
• Training reinforcement learning agents to execute multi-step sequences like 'send email' and 'schedule meeting'.
• Evaluating agent generalization across the five different simulated workplace databases.
• Developing new RL algorithms for tool-use and planning in structured, text-based environments.

Strengths

• Contains 690 distinct tasks for agent evaluation.
• Provides 26 distinct tools for agent interaction.
• Includes five different simulated workplace databases.
• Released by NVIDIA in 2026 as part of the NeMo Gym framework.

Limitations

• Specific dataset size, row count, and file formats are unknown.
• Task complexity and potential success rates are not detailed.
• The license for commercial use is unspecified.

Provenance

Source

NVIDIA

Collection Method

Constructed as a synthetic sandbox environment for agent testing.

Freshness

Last updated on 2026-02-26.