Research that shapes the taxonomy
Seven selected papers connect computer use, organizational work and evaluation. The implications below are our design proposals. This is a focused reading list, not a systematic review or a claim that the benchmarks have been reproduced.
TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks
A simulated company joins software tools, coworker communication and execution-based checks. It motivates measuring work across applications and people, rather than treating each interface action as the unit of progress.
Design implication: CUA-Anything proposal: preserve shared company state and verify handoffs and final artifacts.
Reading scope: Full-text introduction and evaluation sections
Official implementation ↗WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks?
ServiceNow tasks and BrowserGym connect enterprise knowledge work with a reproducible browser-agent environment. Task families, generated instances and software platforms are different counting units.
Design implication: Keep application coverage, task definitions and sampled runs separate.
Reading scope: Full-text abstract and benchmark description
Official implementation ↗WorkArena++: Towards Compositional Planning and Reasoning-based Common Knowledge Work Tasks
Composes atomic actions into knowledge-work workflows requiring planning, retrieval, reasoning and memory. Its randomized instances make a fixed evaluation curriculum important for fair comparisons.
Design implication: Build role episodes from verified subtasks, and pin seeds, environment revisions and evaluation criteria.
Reading scope: Full-text introduction and evaluation curriculum
Official implementation ↗OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments
Pairs real desktop and web tasks with initial-state setup and execution-based evaluators. It supports distinguishing an attractive task idea from a reproducible, verifiable task package.
Design implication: Keep setup, verifier, reset and source provenance in each accepted CUA-Anything package.
Reading scope: Full-text abstract and task/evaluation description
Official implementation ↗WorkWorlds: An Infrastructure for Evaluating AI Agents on Workplace Tasks
Fixes organizational state, date and employee access before specifying tasks. This highlights how hand-picked task context can simplify information discovery and distort workplace evaluations.
Design implication: Test whether a task remains valid in a shared world with irrelevant and changing information. Abstract-level reading; no reproduction performed.
Reading scope: Abstract and version metadata only; preprint
Do's and Don'ts for Human and Digital Worker Integration
Argues for multiple autonomy and human-involvement levels, and measures beyond productivity when integrating digital workers into business processes.
Design implication: Record escalation, human effort and recovery alongside task completion; this is conceptual guidance, not benchmark evidence.
Reading scope: Abstract only; position paper
GEO: Generative Engine Optimization
Studies content visibility in a particular generative-search benchmark, including citations and statistics. Experimental visibility gains do not establish current Google ranking rules or guarantee results for this site.
Design implication: Publish inspectable evidence and precise definitions, then measure real citations and traffic rather than promising a ranking uplift.
Reading scope: Full-text methods and evaluation sections
Search discovery: what the evidence supports
Our approach is readable HTML, stable URLs, original evidence, clear definitions and a sitemap. Google states that its AI search features use core search foundations and do not require special AI text files. OpenAI identifies OAI-SearchBot as its search crawler. IndexNow notifies participating engines of changed URLs; it does not certify indexing.