Work Agents: 16 Researchers Build a Single Coordinate System for AI Agents

A team of 16 researchers has released a survey paper, Work Agents: Systems, Capabilities, and Optimization (arXiv 2610.35845). The authors try to answer a question that concerns everyone deploying agents right now: how do you make sure a delegated task ends in a verifiable result? The authors include Zishan Xu, Enjun Du, Mingqi Gao, Cheng Yang, Junhao Shen, Difan Zou and Weiwen Liu. The paper collected 15 upvotes on Hugging Face Daily Papers, and it comes with a project page and a GitHub repository, awesome-working-agent, that collects the literature.

An Agent Is a System, Not a Model

The starting thesis is that foundation models have moved from generating information to carrying out digital work. Agents now operate across tools, applications and workflows. Yet, in the authors' view, the research lacks a unified picture of how an agent reliably does what it is assigned.

They propose viewing a work agent as a system with two parts: a foundation model and a runtime harness, meaning the software layer in which the model runs. The agent solves tasks in a stateful workspace, a workspace that preserves its state, and must produce a verifiable result at the end. It follows that reliability depends not only on the model weights, but also on how memory, tools, step control and acceptance of the final result are built.

Two

Axes, Five Forms of Work and Five Families of Capabilities

To describe the conditions in which an agent works, the authors introduce two dimensions:

• Work persistence: how long the work runs and how much it relies on state preserved between steps and sessions.
• Work openness: how open and unpredictable the environment is, as opposed to being limited to a set of actions known in advance.

By the scale and organization of tasks, digital work falls into five forms. This classification underpins a list of five families of capabilities needed for reliable execution. Together they provide a shared vocabulary: you can compare an agent built for a one-off operation in a closed application with an agent that runs a project for weeks in an open environment.

Where to Optimize: Model, Harness or Both

A separate part of the survey covers how to improve agents. The authors examine three directions:

• Model optimization: fine-tuning and otherwise improving the model itself.
• Harness optimization: tuning the harness, meaning how the agent is built around the model and how it interacts with the workspace.
• Joint optimization: tuning the model and the harness together.

This split is useful in practice. Teams often credit an agent's progress to a new model when the gain may have come from a rebuilt harness, and vice versa. Optimizing both parts together accounts for their mutual dependencies explicitly.

Evaluating by Work Outcome

The authors analyze evaluation methods and metrics from the standpoint of work outcomes, that is, the results of completed work. What should be measured is not how plausible an answer sounds, but whether the task is solved and whether that can be verified. For business this is fundamental: a customer pays for a finished report, a closed ticket or working code, not for a polished chain of reasoning.

The section on the future names three directions:

• sustained work delegation over long horizons;
• adaptation to open environments;
• the interplay between doing work and the growth of the agent's capabilities.

The last item is interesting because an agent that performs tasks can accumulate experience and improve on its own, and the research agenda treats execution and learning as linked processes.

How to Use the Framework When Choosing an Agent Stack

The survey does not provide a ready-made benchmark, but it does provide a checklist for design and procurement. Before comparing agents on leaderboards, it is worth answering a few questions:

• How long does your work run, and where is its state stored?
• How open is the environment: a fixed set of APIs, or arbitrary applications and websites?
• What in your agent belongs to the model and what to the harness, and which part can you change?
• By what verifiable result do you accept the work?

Teams that answer these questions before choosing a model are more likely to see exactly where an agent's reliability breaks down, and are less likely to look for the cause of failures in the wrong place.