Microsoft's CORPGEN Boosts AI Multitasking

Current AI agents struggle with the complex, interdependent multitasking inherent in real-world corporate environments. While benchmarks typically test one task at a time, a new initiative from Microsoft Research, dubbed CORPGEN, aims to bridge this gap with advanced 'digital employees' designed for genuine workplace productivity.

Traditional agents rapidly degrade under multi-task loads. In Microsoft's Multi-Horizon Task Environments (MHTEs), which simulate dozens of concurrent tasks over several hours, leading systems saw completion rates plummet from 16.7% to just 8.7%. CORPGEN tackles these limitations through hierarchical planning, isolated memory, and adaptive summarization, preventing information overload and cross-task interference.

The CORPGEN framework introduces digital employees with persistent identities and role-specific expertise. These agents operate Microsoft Office applications via GUI automation, maintaining consistent performance across hours of activity. Significantly, CORPGEN achieved up to 3.5 times higher completion rates than baseline agents, completing 15.2% of tasks compared to 4.3% for others under a 46-task load.

Crucially, the largest performance gains stemmed from experiential learning. Agents that store and reuse patterns from completed tasks significantly outperform those treating each task in isolation. This system-level design, rather than reliance on any single base model, ensures CORPGEN's benefits scale directly with underlying model improvements. Digital employees also collaborate via standard channels like email and Microsoft Teams, forming virtual organizations without explicit programming.