Today we’re releasing HelixCove, a dataset of work unfolding across an organization. It captures the trajectories behind workplace outcomes: how information moves between colleagues, how decisions evolve, and how actions change the context for the work that follows.
This first release covers three weeks of work at a fictional biotechnology company: twelve employees, 1,912 work sessions and 59,768 actions across different apps including Word, Outlook, Teams, Excel, SharePoint and more.
We call this data Human Traces. A trace links what someone sees and does to the activity around it: the conversation that prompted an action, the document it changed, and the colleague whose next decision depends on it.
Our thesis is that the most valuable data for AI to do real work is how people actually work. Most workplace data records the outputs of work. Human traces show how those outputs came to exist, through intermediate decisions, incomplete knowledge, corrections and coordination.
HelixCove was generated using patterns from hundreds of thousands of hours of real work captured by Arbium. All people, documents, and events in HelixCove are fictional.
Work as a connected history
A single issue in HelixCove can move between people and applications over several days.
Priya reports a failed reproducibility test in Microsoft Teams: 34% variation against a 20% limit. Daniel replies that no further runs should proceed and saves DEV-26-011 in SharePoint’s QA records, owned by himself. He then asks Priya to lead the investigation, and she accepts. Each step has a screenshot from the captured execution.
Has Daniel done what he said? Who owns the investigation now? These questions have answers only if the connections between events are kept.
What this data enables
Each capture has a timestamp, application and window, event type, and accessibility-tree text, with screen recordings for visual context. Facts planted in the story have known answers, and 430 questions test whether a system can recover them, from simple lookups to changes over time and questions with no answer in the data.
Prior academic work uses computer-use observations for task and skill learning, user modeling and goal inference, and TheAgentCompany and AgentSynth test agents on workplace tasks. HelixCove captures work unfolding across a shared workplace, where one person’s actions shape what others know and do next. This provides a dataset for studying how AI can learn from computer-use observations to complete tasks that depend on colleagues, shared information, and work already in progress.
Generating work in a shared world
We start by specifying a fictional company: its twelve employees, shared documents, workstreams and the events that change the business. Every fact has an availability time and an audience, so one employee can learn of a problem before the others. Story events unfold across the days, connecting facts across people and workstreams.
Agents then play the employees on virtual machines running the real applications in one shared workspace. We capture each action and organize the data by person and time.
Grounding generation in real work
The activity needs the shape of a working day, with bursts of work, returns to earlier applications and variation between roles and days.
We measure eighteen features of that shape, from burstiness and returns to event-type mixtures, text sizes and day-to-day variation.
Real people differ from each other, so we compare against a range. We cut real people’s days into slices, pool them with equal weight per person, and take the 5th to 95th percentile across slices as the range.
We compare both typical activity and less frequent patterns, such as long pauses between actions. For each measure, we check the average, the midpoint and the upper end of the distribution, as well as its overall shape. This helps distinguish histories that look similar on average but behave differently over time.
We then count how many measures fall outside the real-work reference ranges. Real people can fall outside those ranges too, so we test people excluded from the reference group to establish a baseline. The goal is for generated histories to deviate no more than a typical real person does. Comparisons use matching operating systems, capture formats and durations.
Privacy at the source
HelixCove separates the behavioral patterns of real work from the business content of the generated workplace. Real-work observations inform measurements such as activity timing, application transitions and text sizes. The employees, documents, conversations and business events in HelixCove are fictional.
This separation lets us study how work unfolds without making the source workplace’s documents and conversations the dataset.
What comes next
HelixCove is a first release. It covers three weeks of one company’s work, and it is not the whole of what we are building. We are continuing the work on richer world state, more varied workstreams, stronger behavioral calibration and more reliable agents, and on how much surrounding context contributes to learning. It is our first step toward making the full trajectory of work a resource for AI research.