Programmes / Computer Use

Real work, recorded on real software.

 Waitlist 

Agents need to watch people finish work, not read about it.

We are recording task trajectories with operators doing their own jobs: the software they actually use, the detours, the corrections, and the point at which a task is genuinely finished.

Licensing isn't open yet. Join the waitlist and we'll send the first programme's scope and protocol when it's ready.

Built for agent training · agent evaluation · UI grounding

What each trajectory contains.

Sequenced actions

What was done, in what order, with the timing between steps preserved.

Intent and correction

Where the operator changed approach, and why, captured alongside the recording.

Completion criteria

An explicit definition of done per task, so a trajectory can be scored rather than guessed at.

Screen and input capture

Screen recording, clicks, keystrokes and application state, time-aligned.

Join the waitlist

Approach

How computer-use trajectories will be recorded.

01

Real operators, real software

Operators record their own work on the tools they already use. Tasks are not staged in sandboxes built for the recording.

02

Two-sided consent

Recording is agreed with the operator and with the organisation whose software and data appear on screen. Sensitive content is masked to rules set in advance.

03

Annotate intent

Each trajectory is labelled with the task goal, the points where the operator changed approach, and an explicit definition of done.

04

Deliver scored trajectories

Action sequences, screen captures and timing are delivered together with completion criteria, consent records and a processing log.

Tell us what your model is missing.

Scope a collection hello@skulp.ai