'Work Agents' Survey Frames Agents as Model-Plus-Harness Systems
TL;DR
- The survey defines work agents as foundation models paired with runtime harnesses that execute tasks in stateful workspaces and deliver verifiable outcomes.
- Two dimensions, work persistence and work openness, characterize working conditions; digital work is sorted into five forms by task scope and organization.
- The authors identify five core capability families and compare three optimization paths: model, harness, and their joint optimization.
An agent should be understood as a model paired with a runtime harness, a stateful workspace, and the machinery to deliver a verifiable outcome, according to a work-centric survey posted on arXiv on October 26. The paper, 'Work Agents: Systems, Capabilities, and Optimization,' pushes the harness into view as a first-class part of the system rather than scaffolding wrapped around the LLM.
The authors characterize working conditions along 'two complementary dimensions, work persistence and work openness,' and sort digital work into five forms by task scope and organization. From there they define five core capability families required for reliable work execution and examine three optimization paths: model, harness, and their joint optimization. They also pull existing evaluation methods back to the question of work outcomes rather than model behavior in isolation.
The survey lands into a dense flow of agent research on our agents tracker, which now carries 531 stories from the past 90 days. The authors close by flagging three open problems: 'sustained work delegation, adaptation to open environments, and the interaction between work execution and capability improvement.'
Originally reported by arxiv.org
Read the original article →Original headline: 'Work Agents' Survey Reframes Agents as Model-Plus-Harness Systems, Maps Five Work Classes and Joint Optimization