Work Agents: Systems, Capabilities, and Optimization

Work Agents. Systems, Capabilities, and Optimization. By Zishan Xu and fifteen others, from Shanghai Jiao Tong University, the University of Hong Kong, and six other institutions. Published on the twenty-sixth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Foundation models are moving from generating information to executing digital work. Agents built on them are increasingly able to complete complex tasks that cross tools, applications and workflows. What the existing research lacks, the authors argue, is a unified perspective on how such agents reliably accomplish work that has been delegated to them. The emphasis falls on reliability and on delegation. The question is not only whether an agent can do a task once, but whether work handed to it gets done, and whether the result can be checked.

Their answer is a perspective that is centred on work and pitched at the level of the whole system. A work agent, in this view, is a system that combines a foundation model with a runtime harness. Together they execute tasks in stateful workspaces and deliver outcomes that can be verified. Three parts of that definition carry weight. The model is only one component, the harness that runs it is the other. The workspace holds state across the task. And the product is an outcome someone can verify, not just a stream of generated text.

Two complementary dimensions characterise the conditions of work: work persistence and work openness. Digital work itself is classified into five forms, according to the scope of the task and how it is organised. Building on that framework, the survey identifies five core capability families required for reliable execution. It then examines three routes to improvement: optimizing the model, optimizing the harness, and optimizing the two jointly.

Evaluation is read through the same lens. Existing evaluation methods and metrics are analysed from the perspective of work outcomes, which keeps the question tied to whether delegated work was actually delivered.

The future directions follow from the framework as well. They include sustained work delegation, adaptation to open environments, and the interaction between executing work and improving capability. Each one stretches a dimension the survey already named: delegation that lasts longer, environments that are more open, and an agent whose work feeds back into what it can do.

The organising move is to stop treating the model as the whole agent. Once the harness counts as half of the system, and the workspace is understood to carry state, optimization has two places to act, and evaluation has a concrete target: a reliable, verifiable and practically valuable outcome.