Operating a sample page with a browser agent, under control
By Oliver Schmalzl · Written: · Archive discovery date:
10 June 2026 · Archive date
A browser agent can operate an interface instead of just talking about it. For a first attempt, your own test page is particularly clear: you know its content and its expected behaviour.
The task could be to sort three invented paper models by size and to select the smallest one in a local watch list. There are no real accounts, purchases or messages. This keeps the attempt limited to an observable operating task.
Name the goal and the permitted actions
Before the start, you decide which page may be used and what counts as success. An open menu is not yet a saved result. A click on "Select" has to show up in the visible state of the watch list.
Skyvern documents an approach to browser automation. The actual support depends on the tool and the page. For your own attempt, what counts is the sequence you actually observe, not a blanket statement about any website.
If an element cannot be found, the agent should report this or reorient itself within the permitted steps. Using a different page or inventing extra actions would not fulfil the original task.
Recognise completion by the result
After the run, you check: is the right entry selected? Does the sorting have the right meaning? Was anything changed that was not part of the task? Comparing the state before and after the attempt makes these questions tangible.
A good test page can also contain a missing input or a deliberately simple error case. This shows whether the run stops in an understandable way. The exercise stays small, but it gives far more insight than a demonstration without a recognisable limit to the task.
Original finds
These links were the starting point for this post. The assessment in the text is my own; the linked posts are not my own tests.
- Chrome Buddy as a browser agent with a transparent model, your own API access and confirmation of consequential actions.
- Skyvern as an open-source agent for visual web automation.
- Presentation of a fast browser agent called Jev Ultrafast; the post mentions DOM-based operation and the limits of the early project.
- Description of a voice-controlled Mac assistant that is meant to choose actions with Jev; according to the description, screen control and input come from other software.
About the date: This post was written on 11 October 2026. The archive date corresponds to the earliest related find in the collection, not to a publication at that time. Technical assessments and sources added later refer to the time of writing.