TEMPORAL RESPONSIBILITY
Work persistence
From responsibility that closes at delivery to responsibility that remains open across stages, feedback, and later events.
Closes at deliveryStanding obligation
A SURVEY OF DIGITAL KNOWLEDGE WORK 2026
Work Agents:
Systems, Capabilities, and Optimization
1Shanghai Jiao Tong University2The University of Hong Kong3The Hong Kong University of Science and Technology (Guangzhou)4Tsinghua University5The Chinese University of Hong Kong6Zhejiang University7Sun Yat-sen University8Central South University
From generating responses to carrying out digital work.
Five forms of digital work
- 01
Conversation Level
Interpret evidence · Produce answers
- 02
Atomic Action Level
Edit a local object · Verify the change
- 03
Multi-step Level
Dependent actions · Complete an objective
- 04
Workflow Level
Coordinate stages and handoffs
- 05
Workspace Level
Connect related objects · Track dependencies
Expanding work scope and dependencies.Explore the survey
The central thesisFrom answers to
From answers to
accepted work.
The unit of progress is an outcome that meets its requirements and can be used downstream.
Read the introductionAbstract
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.
From the manuscript · Systems, capabilities, optimization, and evaluation
THE WORK LANDSCAPEFigure 4
Figure 4 · Evolution of agent work through 90 representative works in 2023–2026. Colors classify selected task settings.
Explore the figure01 / FoundationTwo dimensions.
Two dimensions.
Different questions.
Persistence asks when delegated responsibility can close. Openness asks what must be discovered during execution. They describe working conditions alongside the scope and structure of work.
THE TASK WORLD
Work openness
From enumerated materials and interfaces to a workspace whose objects, relations, authority, and actors must be discovered while working.
Predefined task worldOpen workspace
These axes describe working conditions. They are independent of intelligence and autonomy.
Six elements, one working system
Task → Execution → Outcome → Verification
A foundation model supplies reasoning and decisions. A harness organizes context, tools, memory, and control. They act in a stateful workspace under task requirements.
(T, E₀) M, H⟶ (τ, Eₜ, O)V(T, E₀, τ, Eₜ, O) ⟶ S
Task & initial state → trajectory, resulting state & outcome → assessment
02 / Work formsWhat work do
What work do
agents undertake?
Five forms describe the scope and organization of delegated work. Scope concerns objectives, stages, objects, and dependencies.
Conversation Level
Produce an informational outcome.
Interpret a request and its associated evidence to produce an answer, analysis, classification, judgment, or specification. Completion is assessed on the returned information, without requiring a change to source artifacts or external state.
Read in the surveyA WORK SCENARIO
Analyze a financial report and answer a numerical question.
Supplied context → Answer or judgment
Representative resourcesDocVQA · FinQA · LegalBench
Atomic Action Level
One local objective, a concrete effect.
Apply a local operation to a clearly defined work object and confirm its intended effect. The scope of the objective defines this form; several mouse, keyboard, or API operations may still be needed.
Read in the surveyA WORK SCENARIO
Correct a formula in cell C7 and confirm the intended result.
Target object → Confirmed local change
Representative resourcesInstructExcel · PPTC · ScreenSpot-Pro
Multi-step Level
Coordinate actions toward a complete result.
Coordinate dependent actions toward a complete objective. Intermediate results supply evidence or state for later decisions, and execution feedback can change the remaining sequence. Completion is assessed on the final result.
Read in the surveyA WORK SCENARIO
Prepare data, run analysis code, and deliver an executable analysis.
Related actions → Complete objective
Representative resourcesScienceAgentBench · DABstep · ToolSandbox
Workflow Level
Keep stages, rules, and handoffs aligned.
Advance a delegated operational process through stages, procedural requirements, approvals, and interactions with other participants. Completion covers the process outcome as well as the final output.
Read in the surveyA WORK SCENARIO
Collect reporting data, prepare artifacts, obtain approval, and publish.
Dependent stages → Completed process
Representative resourcesWorkArena++ · OfficeBench · τ-bench
Workspace Level
Discover the work world as work proceeds.
Establish which files, records, and applications bear on the goal and how they are related. Locate relevant objects, reconcile evidence and versions, and preserve consistency among affected objects when updates are required.
Read in the surveyA WORK SCENARIO
Trace a report to its source workbook and update dependent presentations.
Distributed objects → Consistent work outcomes
Representative resourcesWorkspace-Bench
See 90 representative task settings from the current manuscript
Work forms describe scope and organization. Persistence and openness describe the conditions in which that work is carried out. A longer sequence of actions can still close at delivery.
03 / CapabilitiesWhat makes
What makes
work reliable?
Five capability families form a recurring cycle around shared work context. Evidence guides decisions; verification can send work back to the capability where a problem arises. The running example updates a workbook, report, charts, and slides.
Perception & understanding
Interpret objects and relationships.
Read native work objects, identify relevant evidence, and trace sources, versions, and dependencies across the workspace.
Reasoning & decision making
Make grounded decisions.
Reconcile evidence with goals, professional knowledge, business rules, and constraints before choosing an action.
Planning & coordination
Organize dependent work.
Decompose objectives, track prerequisites, coordinate stages and participants, and manage branches and handoffs.
Execution & state management
Turn intent into effects.
Act on the workspace, distinguish attempted actions from committed changes, and preserve consistency across objects and applications.
Verification, recovery & replanning
Establish and repair the result.
Check outcomes against requirements, investigate uncertain effects, and recover or replan when assumptions fail.
The capabilities form a cycle. An accepted delivery may close a bounded task while an ongoing responsibility remains active.
Explore the five-capability cycle
04 / OptimizationImprove the model.
Improve the model.
Improve the harness.
Diagnose the failure before choosing what to update. The right intervention can reside in model parameters, external support, or their coupling.
Where agents practice
Work materials, services, dependencies, and situations in which actions have observable effects.
What agents learn from
Domain corpora, demonstrations, interaction trajectories, and examples of recovery.
What guides improvement
State checks, outcome verification, rubrics, and rewards that connect behavior to requirements.
Work execution→Observable experience→Diagnosis & feedback→System update ↺
Learn the
recurring capability.
Internalize reusable interpretation and decision patterns in model parameters.
- Pre-training: knowledge & representations
- Mid-training: interaction & long context
- Post-training: SFT, RL & distillation
Improve the
external support.
Repair what the model can observe, invoke, remember, and verify during execution.
- Observation & action interfaces
- Workflows & execution control
- Long context, memory & skills
Adapt the
working pair.
Account for the interaction between learned behavior and the harness supporting it.
- Harness-aware policy training
- Joint model–harness adaptation
- Self-improvement & recursive improvement
Compare (M₀, H₀), (M₁, H₀), (M₀, H₁), and (M₁, H₁) on matched tasks, states, verifiers, and budgets to separate model gains, harness gains, and their interaction.
See how experience becomes system capability
05 / EvaluationWhat does the
What does the
evidence establish?
Connect the work requested to the evidence inspected, the judging method, and the reported metric. Assess quality and reliability alongside cost and value in use.
01Requested workGoals & constraints
→02Observed evidenceAnswer, state, artifact, trace
→03Judging procedureRules, rubrics, models, experts
→04Reported metricSuccess, quality, reliability, cost
→05Capability claimBounded by inspected properties
Answer
Does the response satisfy the question and agree with its supporting evidence?
Reference answers · Source supportState
Did the intended effect actually happen, with relevant constraints and other state preserved?
Tests · Database checks · EffectsArtifact
Is the native deliverable correct, editable, source-consistent, and usable in its intended form?
Native objects · Rendered viewsWorkflow
Were dependencies, prerequisites, approvals, and intermediate obligations respected?
Checkpoints · Traces · HandoffsMatch evidence to the requirement
A task can request a correct, editable, source-consistent workbook while a verifier checks selected cells. Those checks establish correctness over the inspected cells; editability, provenance, and other structure need corresponding evidence.
BEYOND A SUCCESS SCORE
Accepted work meets the brief. Value in use concerns what the result helps people achieve. Total effort includes human supervision, correction, and computational cost.
Performance, cost & value ↗The reading roomExplore the
Explore the
research landscape.
30 representative benchmarks and systems drawn from the survey. Browse by work form, then compare the work domain and evidence each resource inspects.
30 resources
| Resource / year | Work form | Domain | Assessment evidence | Paper link |
|---|---|---|---|---|
| Workspace-Bench2026 · Ref. 36 | Workspace | Professional | Rubrics + file-dependency graph | ↗ |
| OfficeBench2024 · Ref. 76 | Workflow | Office | Office environment task completion | ↗ |
| SWE-bench2023 · Ref. 16 | Workspace | Software | Required fixes + regression tests | ↗ |
| OSWorld2024 · Ref. 7 | Multi-step | Computer use | Task-specific execution checks | ↗ |
| AppWorld2024 · Ref. 37 | Multi-step | Enterprise | Goal state + collateral-state checks | ↗ |
| GDPval2025 · Ref. 279 | Multi-step | Professional | Blinded occupational-expert comparison | ↗ |
| PresentBench2026 · Ref. 17 | Multi-step | Office | Model-judged binary checklist on rendered decks | ↗ |
| τ-bench2024 · Ref. 29 | Workflow | Enterprise | Database goal-state comparison | ↗ |
| τ²-bench2025 · Ref. 30 | Workflow | Enterprise | Shared-state and policy evaluation | ↗ |
| TheAgentCompany2024 · Ref. 31 | Workflow | Enterprise | Intermediate checkpoints + final completion | ↗ |
| WorkArena++2024 · Ref. 133 | Workflow | Enterprise | Enterprise application outcomes | ↗ |
| CRMArena2024 · Ref. 148 | Workspace | Enterprise | Answers to CRM queries | ↗ |
| CRMArena-Pro2025 · Ref. 302 | Workspace | Enterprise | Answer correctness + confidentiality awareness | ↗ |
| SpreadsheetBench 22026 · Ref. 121 | Multi-step | Office | End-to-end spreadsheet workflow assessment | ↗ |
| EnterpriseOps-Gym2026 · Ref. 139 | Workflow | Enterprise | SQL outcome + integrity checks | ↗ |
| StartupBench2026 · Ref. 280 | Workspace | Professional | Weighted, tool-assisted rubric judgments | ↗ |
| DABstep2025 · Ref. 126 | Multi-step | Professional | Multi-step data analysis outcomes | ↗ |
| DSBench2024 · Ref. 286 | Multi-step | Professional | Data-analysis and modeling results | ↗ |
| ScienceAgentBench2024 · Ref. 10 | Multi-step | Science | Data-driven scientific task assessment | ↗ |
| PaperBench2025 · Ref. 136 | Workflow | Science | Hierarchical research-replication rubric | ↗ |
| WebArena2023 · Ref. 6 | Multi-step | Computer use | Answer or required website state | ↗ |
| SpreadsheetBench2024 · Ref. 56 | Multi-step | Office | Target cells across workbook test cases | ↗ |
| ToolSandbox2024 · Ref. 127 | Multi-step | Enterprise | State milestones + specified minefields | ↗ |
| SheetCopilot2023 · Ref. 4 | Multi-step | Office | Spreadsheet operations and task outcomes | ↗ |
| InstructExcel2023 · Ref. 111 | Atomic Action | Office | Concrete Excel instruction execution | ↗ |
| API-Bank2023 · Ref. 25 | Multi-step | Tool use | API selection, arguments, and execution | ↗ |
| ScreenSpot-Pro2025 · Ref. 115 | Atomic Action | Computer use | Localized interface target grounding | ↗ |
| DocVQA2020 · Ref. 22 | Conversation | Professional | ANLS against reference answers | ↗ |
| FinQA2021 · Ref. 21 | Conversation | Professional | Numerical result + program equivalence | ↗ |
| LegalBench2023 · Ref. 101 | Conversation | Professional | Expert-designed labels or references | ↗ |
No matching resources.
Try another keyword, work form, or domain.
Representative placements are a reading aid; work forms can overlap. Reference numbers follow the current manuscript; years identify the selected releases. Evidence descriptions summarize the stated configurations.
See the paper’s domain map across 79 benchmarks
06 / The path forwardBeyond bounded tasks.
Beyond bounded tasks.
Toward dependable work.
Personal work agents, human digital twins, and organizational work systems make sustained delegation concrete. They motivate advances in environment, persistence, and verification.
Personal work agents
Carry a user’s goals across requests; advance unfinished work and choose when their attention is needed.
Human digital twins
Represent a particular person’s judgment, with separate checks for professional quality and personal fidelity.
Organizational work systems
Connect contributions, decision rights, and dependencies across people and agents.
Environment
Generalization needs realistic work.
Construct work environments that preserve realistic objects, relationships, histories, and professional requirements. Study transfer across work settings and differences between people.
Held-out transfer · Source selection · Realistic curriculaPersistence
Responsibility outlasts one execution.
Retain goals, judgments, commitments, and dependencies across events and interruptions. Choose when to observe, ask, act, or wait, and adapt procedures without losing track of current work.
Fulfilled obligations · Duplicate effects · ContinuityVerification
Assess quality, fidelity, and value.
Assess deliverable quality, a digital twin’s agreement with the represented person, and the return on delegation separately. Follow benefits and human and computational effort through delivery and downstream use.
Work quality · Personal fidelity · Return on delegationExplore the next-generation research agenda
Two interacting loopsDeliver today.
Deliver today.
Improve the next cycle.
Trace an update’s effects while current responsibilities continue.
Carry current responsibilities
Observe & decide → Act & verify → Update work state
Retain agreed requirements, source versions, and pending work.Learn for later work
Diagnose → Propose a scoped update → Test independently
Use execution history and feedback from downstream use.Before applying an update: trace affected results and commitments. Weigh capability gains against rechecking, repair, and coordination costs, then apply the update, migrate affected work, or retain the current procedure.
Explore how updates return to ongoing work
The measure of progress
Work delivered.
Outcomes accepted.
Read, explore, referenceContinue the
Continue the
conversation.
The full survey connects the foundations, capabilities, optimization methods, evaluation resources, and open challenges of work agents.
Read the full surveyBibTeX arXiv:2610.35845 · 2026
@misc{workagents2026,
author = {Zishan Xu and Enjun Du and Mingqi Gao and
Cheng Yang and Junhao Shen and Zirong Chen and
Zhiyuan Yao and Yifu Guo and Yuquan Lu and
Jinyang Huang and Zhengxi Lu and Siyi Liu and
Ziyu Zheng and Yongqi Zhang and Difan Zou and
Weiwen Liu},
title = {Work Agents: Systems, Capabilities,
and Optimization},
year = {2026},
eprint = {2610.35845},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2610.35845}
}