Work Agents: Systems, Capabilities, and Optimization

TL;DR

The paper introduces a system-level framework for work agents that treats them as integrated systems combining foundation models with runtime harnesses to perform tasks in persistent, stateful environments. It defines two key dimensions—work persistence and work openness—to classify digital work and identifies five capability families essential for reliable task execution. The framework also guides optimization strategies and evaluation methods focused on verifiable outcomes.

Why It Matters

This framework enables more reliable and scalable deployment of AI agents in real-world workflows by providing a structured approach to design, optimize, and assess their performance. It supports the transition from AI models that generate content to those that execute complex, multi-step tasks with accountability.

Key Findings

• Work agents are best understood as systems integrating foundation models with runtime harnesses operating in stateful workspaces.
• Work persistence and work openness serve as complementary dimensions for characterizing the conditions under which digital work occurs.
• Five distinct forms of digital work emerge based on task scope and organization, each requiring specific capability families for reliable execution.