Product · Agent
Perceive, plan, act, verify.
Every task runs the same loop. The agent reads the live page, decides one next step, acts on it, then checks its own work before moving on, grounded in what is actually on screen.
One step at a time
The loop runs until the task is done.
Perceive
The agent takes a visual snapshot, reads the rendered DOM, and builds a structured outline of what is on screen.
Plan
Given the task and the current page state, it decides the single next action worth taking.
Act
It executes that action on the live page: click, type, navigate, scroll, extract, or wait.
Verify
It checks that the action had the effect it expected before it commits to the next step.
Three ways of reading one page.
The agent never guesses at layout. It combines structure, pixels, and semantics into one grounded picture before it decides anything.
Rendered DOM
The full document with every interactive element identified and addressable.
Visual snapshot
A screenshot of the page as rendered, used for spatial reasoning.
Structured outline
A semantic summary: headings, forms, buttons, links, and tables.
The actions it takes on your behalf.
Navigate and click
Open URLs, click buttons, links, and interactive elements.
Type and fill
Enter text into forms, search bars, and fields.
Scroll and read
Move through content and extract what is visible.
Download
Save files, with explicit approval first.
Upload
Attach files to forms and upload interfaces.
Multi-tab
Coordinate work across several tabs at once.
Grounded, not hallucinated.
Every action selects from elements that actually exist in the DOM. There are no fabricated clicks and no imagined buttons. When the page changes, the verify step catches it, and the loop replans against the page as it now is.