A framework connects digital work agents to the work they must finish
Work Agents: Systems, Capabilities, and Optimization · arXiv:2610.35845
In plain words
This paper is a review and framework for understanding software agents that do delegated digital work. It looks at the model making decisions, the supporting software that manages tools and work, and the digital workspace where files or records change. It asks what work must be done and how completion is checked.
The framework separates when responsibility ends from how much the agent must discover along the way. Its review of 79 representative tests also shows that a score means only what the test’s checks cover.
The research
When an agent works across tools and applications, a correct final reply may not show whether it changed the right files, completed the process, or met the task’s constraints. How can research describe reliability across the whole assignment, rather than infer it from a reply or one action?
The paper connects agent components, work conditions, and outcome checks in a framework for understanding delegated digital work.
The review asks how to understand reliable delegated work when tasks span tools, applications, and workflows. It defines a work agent as a model and supporting software acting in a digital workspace, with outcomes checked against task requirements.
It then relates work conditions and scope to the capabilities needed to carry tasks through. A review of representative tests shows that evaluation may inspect different results and requirements, so a score supports conclusions only about the checks it uses.
This is a conceptual review and framework, not a controlled comparison of agent systems. Its focus is digital knowledge work in the named fields, and it does not establish how well the framework applies beyond them.
What the evidence shows
- Start with completed work
The paper presents a work-centered framework and review for agents that carry out delegated digital work. It starts from a gap in existing research: studying model outputs or individual tool calls can miss how task requirements, workspace state, execution, outcomes, and checks fit together. The paper therefore asks readers to consider whether the complete work meets its requirements.
- Define the system
The framework defines a work agent as a foundation model making decisions together with a runtime harness that organizes context, tools, memory, and execution. They act in a stateful digital workspace under a task’s constraints and permissions.
The result is assessed against task requirements using observable evidence, so the object of study includes what changed and whether the requested work was completed.
- Separate conditions and scope
Two separate dimensions describe working conditions. Persistence concerns whether responsibility ends at delivery or stays open across stages and events; openness concerns whether relevant objects and rules are specified beforehand or must be discovered during execution.
The paper also classifies work by scope, from an informational result or local change to workflows and work across related objects.
- Connect work to capabilities
The framework groups work capabilities into understanding, decision making, planning and coordination, execution and state management, and verification with recovery or replanning. These are functions of the work process, not claims that any particular agent performs them reliably.
The model, harness, or their interaction may carry them out as observations and checks guide further action.
- Read scores in context
The review maps 79 representative tests by their main work domain or setting and distinguishes what evaluation inspects, what it judges, how judgments are made, and how they become scores. Because tests can inspect different results and requirements, their scores do not all measure the same thing.
The framework supports interpreting evaluation in context; it does not show that one optimization strategy is best.
Deep
- Original paper authors
- Zishan Xu, Enjun Du, Mingqi Gao, Cheng Yang, Junhao Shen
- Submitted to arXiv
- License
- CC0