From AI Agents to Cyber Employees: Unveiling the Emergence of Productivity Intelligence

Recent AI developments have shifted from asking whether a model can answer questions to whether an agent can fully automate a workspace, understand personalized needs, and complete tasks end‑to‑end like a human worker. The core question is whether AI can autonomously take over a job, grasp context, and deliver reliable results.

Analysis of over 100 internal Feishu cases shows that current agents operate mainly at the "action layer"—they can write text or open files but struggle to comprehend the broader work environment. The key challenge is enabling an agent to identify which documents to consult, what information to trust, and how to organize outputs into a verifiable deliverable.

The authors introduce the concept of a "cyber employee": an AI unit that possesses its own workspace, understands role responsibilities, autonomously explores task goals, continuously learns, and delivers results that can be validated. This notion underpins the Workspace‑Bench benchmark, which evaluates agents in realistic office settings rather than clean, single‑file demos.

Workspace‑Bench 1.0 constructs five real‑world roles (operations manager, logistics manager, product manager, backend developer, researcher) across a workspace containing 20,476 files, 74 file types, 3,299 directories, a maximum depth of 8, and up to 11,020 files per workstation. The benchmark defines 388 tasks with file‑dependency graphs and 7,399 fine‑grained rubrics, requiring agents to resolve an average of 5.1 dependency edges, span 4.7 files, and satisfy 19.1 evaluation criteria per task.

One representative task asks an operations manager to produce a global product‑strategy report by aggregating data from nine core files (CSV orders, PDF logistics, Excel product info, etc.) and passing 25 rubric checks that assess both the correctness of the result and the process used to obtain it. This mirrors a real‑world “digital office trial” where the agent must reconstruct workflow, synthesize evidence, and generate a deliverable.

Experimental results show a substantial performance gap: overall pass rates on Workspace‑Bench‑Lite range from 27 % to 60 % (average 45.1 %), far below the 80.7 % achieved by human experts with tools. Across 27 agent‑harness and foundation‑model combinations, the average rubric pass rate is 43.3 %, with the best configuration approaching 60 %.

Performance degrades as task difficulty increases: Easy tasks achieve 51.4 % pass, Medium 46.0 %, and Hard only 35.7 %. Hard tasks require discovering file relationships, long‑term planning, state tracking, and error recovery, revealing that agents often get lost in complex dependency networks.

Analysis of dependency‑graph detection shows agents have higher Node F1 than Edge F1, indicating they can locate relevant files but struggle to infer the correct relationships among them—e.g., distinguishing source data from derived reports or outdated versions.

The authors outline three scaling obstacles beyond model size: (1) the sheer scale and heterogeneity of real workspaces; (2) the need to provide a diverse set of role‑specific capabilities; and (3) the breadth of typical productivity tasks that require stable, context‑aware execution. Unlike model scaling, these factors cannot be addressed by larger parameters alone.

Productivity‑intelligence emergence is defined as the point where models, agent harnesses, workspace structures, role contexts, task feedback, and organizational processes form a closed loop, enabling stable, reusable, and scalable delivery in real work. The authors argue that emergence is driven not only by larger models but also by the synergy of harness design, workspace understanding, and role‑specific engineering.

In conclusion, the next AI competition will focus on building the infrastructure for productivity intelligence: scaling workspaces, scaling role coverage, and scaling typical enterprise tasks. Moving from "ability‑centric" AI products to "work‑centric" systems will require agents that can transform high‑level goals into reliable outcomes, akin to human employees who proactively gather data, verify context, anticipate risks, and own the final deliverable.