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

A team of 16 researchers has released the survey Work Agents: Systems, Capabilities, and Optimization (arXiv 2610.35845). The authors try to answer a question that's now top of mind for anyone deploying agents: how do you make sure a delegated task actually gets carried through to a verifiable result? The author list includes Zishan Xu, Enjun Du, Mingqi Gao, Cheng Yang, Junhao Shen, Difan Zou, and Weiwen Liu. The paper picked up 15 upvotes on Hugging Face Daily Papers and comes with a project page and a GitHub repository, awesome-working-agent, collecting related literature.

An Agent as a System, Not a Model

The starting premise is that foundation models have moved beyond generating information to actually executing digital work. Agents now operate across tools, applications, and workflows. Yet, the authors argue, research still lacks a unified view of how an agent reliably completes what it's asked to do.

They propose treating a work agent as a system made of two parts: a foundation model and a runtime harness — the software scaffolding in which the model operates. The agent solves tasks inside a stateful workspace, a working environment with persistent state, and must produce a verifiable result at the end. Reliability, in other words, depends not just on the model's weights but on how memory, tools, step-level control, and output acceptance are all structured.

Two

Axes, Five Forms of Work, and Five Families of Capabilities

To describe the conditions an agent operates under, the authors introduce two dimensions:

• Work persistence — how long-running the work is and how much it relies on state preserved across steps and sessions.
• Work openness — how open and unpredictable the environment is, rather than being constrained to a predefined set of actions.

Based on the scope and structure of tasks, digital work is split into five forms. This classification underpins a list of five families of capabilities needed for reliable execution. Together they establish a shared vocabulary: you can now compare an agent handling a one-off operation in a closed app with one that manages a project over weeks in an open environment.

Where to Optimize: the Model, the Harness, or Both

A separate section of the survey covers how to improve agents. The authors break this into three directions:

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

This breakdown is practically useful. Teams often credit an agent's progress to a new model, when the gain actually came from a redesigned harness, or vice versa. Optimizing both together makes their interdependencies explicit.

Measuring by Work Outcomes

The authors analyze evaluation methods and metrics from the standpoint of work outcomes — the results of completed work. What should be measured isn't how plausible an answer sounds, but whether the task was actually solved and whether that can be verified. This matters for business: a client pays for a finished report, a closed ticket, or working code — not for an elegant chain of reasoning.

The section on future directions names three areas:

• sustained work delegation over long horizons;
• adaptation to open environments;
• the interplay between performing work and growing an agent's capabilities.

The last point is interesting because an agent, by doing the work, can accumulate experience and improve on its own — framing execution and learning as linked processes on the research agenda.

How to Use the Framework When Choosing an Agent Stack

The survey doesn't offer a ready-made benchmark, but it does provide a checklist for design and procurement decisions. Before comparing agents by leaderboard scores, it's worth answering a few questions:

• How long-running is your work, and where is its state stored?
• How open is the environment: a fixed set of APIs, or arbitrary apps and websites?
• What in your agent belongs to the model, and what to the harness — and which part can you actually change?
• What verifiable result do you use to accept the work as done?

Teams that answer these questions before picking a model are more likely to spot exactly where an agent's reliability breaks down — instead of looking for the cause in the wrong place.