Start from
How it works
An environment exposes a browser over the Chrome DevTools Protocol, a desktop screen over VNC, or both. HUD’s built-in agents open the ones they support and give the model its provider’s tools for them, so tasks don’t change from one model to the next. See capabilities.1
Deploy the environment
Deploy your environment with
hud deploy, or use a featured one. See
environments.2
Create tasks
Each template defines a kind of task; each task fills in its arguments, such as the site to
start on and what to find. See create tasks.
3
Run them on several models
Start a job from the taskset and compare models on the same tasks. See
evaluations. A trace holds at most 500 screenshots, so keep
Max steps at 500 or below when the agent takes one per step. See
trace limits.
4
Watch the runs
A trace’s Screen view shows the screenshots from its browser and computer-use steps, and
the conversation shows the action behind each one. See
inspect a trace.
5
Review them with QA
QA agents view each screenshot, so they catch answers that the screens don’t support and
failures caused by the page rather than the agent. See QA agents.
Grade what the agent achieved
Grade the outcome, not the clicks. Have the template ask for a structured answer and check it against the source, or check the state the agent left behind, such as a submitted form or a saved file. The browser-use template gives full reward only when the chair count is right and the screenshots the answer cites show the kitchen, so a correct guess without evidence scores nothing.
Traces
Replay every step and screenshot.
Capabilities
Browser, screen, shell, and tool protocols.
Run QA
Review one trace or every new run.