Skip to content
All products

Agent environments

Computer & browser use training data.

Desktop and browser tasks with expert trajectories for navigation, form completion, document handling, and multi-step workflows.

Discuss your use case

Connect interface actions to completed work.

Browser and desktop work requires an agent to interpret the interface, choose an action, and confirm the result. A completed navigation step can still lead to the wrong record or an unsaved change.

We pair realistic environments with expert trajectories. Each run connects the starting conditions, observations, actions, and final outcome so reviewers can assess the whole workflow.

Capture context alongside actions

Capture the visible state and resulting changes alongside each action. Include the context a reviewer needs to judge whether the next step fits the task.

  • Pair interface observations with action sequences.
  • Document the task and its initial conditions.
  • Retain the artifacts or state needed to verify completion.

Evaluate completion across the workflow

Check the final state and the decisions that affected it. Include recovery examples for incorrect values, failed interactions, and missing confirmations.

  • Check task completion rather than click count.
  • Include navigation errors and recovery paths.
  • Separate training workflows from evaluation cases.

What a program can include

Scope can include:

  • Scoped browser and desktop task definitions
  • Expert observation and action trajectories
  • Starting-state and completion specifications
  • Review labels for navigation, execution, and recovery

Illustrative task / Not a production record

Make the objective concrete.

Objective
Update a project status and attach the correct supporting file.
Context
A representative application, project records, reference files, and a task brief.
Review
Verify that the correct project is selected, fields are changed accurately, and the attachment and update are saved.

Common questions

Why do starting conditions matter?

A documented starting state makes a run understandable and helps teams compare attempts under the same conditions. It also makes it easier to distinguish an execution error from a difference in available context.

Can browser workflows include tool calls?

A scoped workflow can combine interface interaction with API or tool use. Define the boundaries and available actions so that review criteria cover the whole task.

Built around your model

What does your model need to do?

Share representative tasks and examples of a good result. We’ll scope the data and evaluation with you.

Discuss your use case