What is an agent you can hand work to, and how should it be measured? ── A preprint survey that treats work agents as foundation models combined with runtime harnesses and classifies work along persistence and openness

1. What problem does the paper address?

The authors' complaint is that existing research lacks a unified perspective on how agents reliably accomplish delegated work^[1]. Agents are increasingly able to complete complex tasks across tools, applications and workflows. Reports that an agent solved individual tasks, however, make it hard to judge whether real work can be entrusted to it.

Surveys of agents in general already exist. One well-known example organizes the research into construction, application and evaluation^[3]. This paper moves the question from what an agent is to how work gets completed. It is a preprint that has not been peer reviewed, and it is a survey, not an original paper presenting new experimental results. What follows should be read as the authors' organization of the field and as their claims.

2. What does it propose?

The central idea is a definition: a work agent is a system combining a foundation model and a runtime harness. Such a system executes tasks in stateful workspaces and delivers verifiable outcomes^[1]. The key point of this definition is that performance is not credited to the model alone. The execution machinery around the model also becomes an object of analysis.

Several further layers of organization rest on this definition.

• Two complementary dimensions, work persistence and work openness, characterize working conditions.
• Digital work is classified into five forms according to task scope and organization.
• Five core capability families are identified as necessary for reliable execution.
• Means of improvement are split into model optimization, harness optimization and their joint optimization.

3. What did it show? (Only what the abstract states)

Because this is a survey, the paper does not establish new facts through its own experiments. The abstract describes a framework and a scope of review^[1]. Beyond the framework above, the authors analyze existing evaluation methods and metrics from the perspective of work outcomes. As future directions, they name sustained work delegation, adaptation to open environments, and the interaction between work execution and capability improvement.

The abstract also leaves a good deal unsaid. It does not name the five forms or the five capability families, and it does not define the two dimensions. Nor can a reader tell from the abstract what weaknesses the authors found in particular evaluation methods. The "reliable, verifiable, and practically valuable" outcomes they mention describe a goal the framework supports, not a reported achievement.

The companion repository publishes a reading list divided into research on methods and systems and research on benchmarks and evaluation^[2]. It is not the framework itself, but it gives readers a way to check what the survey covered.

4. A concrete example

Because the abstract does not define persistence and openness, this example reads the words in their ordinary sense: how far work carries state forward over time, and how far its surrounding environment is left undetermined. This reading belongs to this article and is not the paper's definition. Consider two cases of asking an agent to revise an internal procedure document.

5. What is not new, and where are the limits?

The individual elements are not new. SWE-bench measures success on realistic tasks by asking whether models can resolve GitHub issues^[4]. GDPval evaluates the quality of deliverables based on the work of industry professionals^[5]. The newer element in this paper is the attempt to line these efforts up from the single viewpoint of completed work.

Several limits apply when reading it.

• There is no primary finding. The value of an overview depends on how its literature was selected and how soundly it was classified, and neither can be assessed from the abstract.
• The framework's usefulness has not been demonstrated. Within the abstract, nothing tests whether the two dimensions and the classification predict an agent's success or failure, or change any design decision.
• "Unified" is the authors' claim. Other ways of organizing the same area exist, and whether this one becomes standard remains to be seen.
• "Verifiable" is left open. The abstract does not say who checks the outcomes or against what criteria.

6. Why might this topic be clustering now?

In the selection record, this paper does not form a cluster. The cluster size is 1 and no related papers are listed. The signals that fired were votes from readers (119 upvotes) and from implementers (70 GitHub stars). These measure attention, and they show neither that the framework is correct nor that the topic is important.

The stars need particular care. The companion repository is a list of papers, so its stars do not indicate that anyone has begun reproducing an implementation^[2]. People may star a list simply as a reminder to read it later.

The keywords include runtime harnesses, stateful workspaces, work persistence and work outcomes. These terms suggest an interest in execution machinery and delivered results more than in the performance of a model on its own. This site has covered other preprints on harnesses, but today's selection data offers no basis for concluding that the direction of research is shifting that way.

7. What connects to pharmaceutical and regulatory practice?

The connection for pharmaceutical document work is the widening of what must be checked, from the model to the whole system. If an agent's behaviour depends on the combination of model and harness, then changing how memory is handled or how tools are called may call for an impact assessment even when the model stays the same. That gives companies one way to think about how far change control should reach.

The idea of verifiable outcomes is also close to everyday practice. In document work, completion should be judged by checking the delivered documents and change records against set criteria, not by the agent's report that it has finished. Viewing the workspace as stateful adds another check: whether changes made during the work can be traced.

However, the paper addresses neither pharmaceutical work nor regulatory requirements. The abstract does not support the claim that following this framework raises quality or brings regulatory compliance. What it offers is vocabulary for separating the conditions of delegated work and deciding how each kind of job should be checked.

When handing work to an agent, the question is more than which model to use. It is what conditions the work runs under, what system runs it, and who checks the delivered result and how.