Work Agents: Systems, Capabilities, and Optimization

AI summary

Overview

Research area: Machine learning / AI agents — specifically the design, optimization, and evaluation of agent systems that carry out delegated digital work. Technical level: Intermediate — the abstract assumes familiarity with foundation models and agent tooling, but the framing is conceptual rather than mathematical. Scope: A work-centric, system-level framework that characterizes how agents combining foundation models with runtime harnesses execute delegated tasks in stateful workspaces and produce verifiable outcomes, along with the capabilities, optimization strategies, and evaluation approaches involved.

What This Paper Is About

Foundation models have moved beyond producing information toward actually carrying out digital work, and agents built on them can now complete multi-step tasks spanning tools, applications, and workflows. Research on these agents is scattered, though, and there is no shared way of describing how an agent reliably completes work that has been handed to it. This paper proposes a unified framework for understanding, optimizing, and evaluating such "work agents" so that the field can reason about them as complete systems rather than as isolated models or isolated engineering tricks.

Key Contributions

1. A work-centric, system-level definition of work agents. The paper frames agents as systems that pair a foundation model with a runtime harness, operating inside stateful workspaces and judged by whether they deliver verifiable outcomes — rather than as models alone or as collections of features.
2. A two-dimensional characterization of working conditions plus a taxonomy of digital work. Two complementary dimensions, work persistence and work openness, are introduced to describe the conditions under which an agent operates, and digital work is classified into five forms according to task scope and organization. (The abstract does not name the five forms.)
3. Identification of five core capability families and three optimization routes. The paper names five capability families needed for reliable work execution and analyzes model optimization, harness optimization, and joint optimization of the two. (The abstract does not enumerate the capability families.)
4. A work-outcome-oriented review of evaluation, plus a research agenda. Existing evaluation methods and metrics are analyzed through the lens of work outcomes, and future directions are laid out covering sustained delegation, adaptation to open environments, and the interplay between executing work and improving capability.

Main Findings

• The framing gap is the central problem: the abstract argues that existing research lacks a unified perspective on how agents reliably accomplish delegated work, and positions the framework as the remedy.
• Persistence and openness are the two axes that matter: working conditions are characterized by how long or sustained the work is (persistence) and how unconstrained or open-ended it is (openness), described as complementary dimensions.
• Digital work is not monolithic: the paper argues it falls into five distinct forms defined by task scope and organization, implying different requirements per form.
• Reliability rests on capability families, not a single skill: five capability families are identified as prerequisites for reliable execution; the abstract does not list or rank them.
• Optimization is two-sided and jointly coupled: the paper treats model-side optimization, harness-side optimization, and their joint optimization as linked concerns rather than independent levers.
• Evaluation should be anchored in outcomes: existing methods and metrics are analyzed from the perspective of the work actually produced, in line with the framework's emphasis on verifiable outcomes.
• No empirical results are reported in the abstract. It describes a framework and agenda; it gives no benchmarks, scores, dataset sizes, or ablations, and none are summarized here.

Methodology in Plain English

This is a conceptual and organizational paper rather than an experimental one. The authors step back from individual agent systems and ask what they all have in common when they do real work: a model doing the reasoning, some surrounding runtime scaffolding that lets it act, a workspace whose state persists across steps, and an outcome someone can check. From that starting point they build a vocabulary — two dimensions describing how demanding the working conditions are, and five categories of digital work — then use it to organize which capabilities an agent needs, where optimization effort can be applied (to the model, to the scaffolding, or to both together), and how success should be measured. Existing evaluation practice is then reviewed against that vocabulary, and the gaps are turned into a set of future research questions.

Why This Matters

Impact on research: The paper aims to give a shared vocabulary and structure to a fast-moving but fragmented area, which matters for comparing systems, deciding where to invest engineering effort, and designing evaluations that reflect real work rather than narrow task scores. By insisting on verifiable outcomes and stateful workspaces, it pushes the field toward measuring delivery rather than generation.

Real-world applications:

• Delegating multi-step business workflows that span several tools and applications, where the agent must keep state and finish the job rather than draft a plan.
• Long-running delegated tasks that need to be sustained over extended periods, which is exactly what the persistence dimension targets.
• Operating in open, unconstrained environments where the task boundary is not fully specified in advance.
• Auditing and quality-checking agent output, since the framework's emphasis on verifiable outcomes implies outcomes that can be confirmed after the fact.

Industry relevance: Teams building or deploying agent products need to decide whether to improve the underlying model, the runtime scaffolding around it, or both — the paper's three-way optimization framing maps directly onto that build-versus-buy-versus-tune decision, and its outcome-based evaluation lens speaks to how deployments get judged in practice.

Future Directions

• Sustained work delegation: how agents can be trusted with work that continues over long horizons rather than a single bounded request.
• Adaptation to open environments: how agents cope when the workspace, tools, or task boundaries are not fixed in advance.
• The feedback loop between execution and capability: how carrying out work changes the agent's own capabilities, treated as an interaction to be studied rather than a side effect.
• Operationalizing the framework: the abstract does not specify how the five work forms or five capability families should be measured, leaving open how the taxonomy and evaluation lens get turned into concrete benchmarks.

Target Audience

Researchers and graduate students working on LLM-based agents and agent architectures; system builders and engineers who design agent runtimes and harnesses; and technical product or evaluation leads who need a structured way to reason about what makes an agent reliable at delegated work. Readers looking for empirical benchmarks or head-to-head system comparisons will not find them in this paper, which is positioned as a framework and research agenda.

Authors’ abstract

As foundation models evolve from information generation to digital work execution, agents are increasingly capable of completing complex tasks across tools, applications, and workflows. However, existing research lacks a unified perspective on how agents reliably accomplish delegated work. To address this gap, we present a work-centric, system-level perspective that views work agents as systems combining foundation models and runtime harnesses to execute tasks in stateful workspaces and deliver verifiable outcomes. We introduce two complementary dimensions, work persistence and work openness, to characterize working conditions, and classify digital work into five forms according to task scope and organization. Building on this framework, we identify five core capability families required for reliable work execution and examine approaches to model optimization, harness optimization, and their joint optimization. We further analyze existing evaluation methods and metrics from the perspective of work outcomes. Finally, we discuss future research directions, including sustained work delegation, adaptation to open environments, and the interaction between work execution and capability improvement. This paper provides a unified framework for understanding, optimizing, and evaluating work agents, supporting the development of systems capable of delivering reliable, verifiable, and practically valuable work outcomes.