fastbrowse
A fast browser agent: Jev picks each action from what is on the page, an LLM reads and plans, and every claim in an answer cites a quote from the page.
A fast browser agent: Jev picks each action from what is on the page, an LLM reads and plans, and every claim in an answer cites a quote from the page.
When the page exposes more controls than fit in the next request, fastbrowse builds state from the task, subgoal, requirements, recent actions, and page identity. Jev evaluates whether each candidate control is relevant; code uses those Noul probabilities to shortlist controls before asking for the next action.
Is this candidate page element relevant to the next actions or information needed for the task? The element’s label, role, context, and link are included in the question.
A task and subgoal, outstanding information requirements, recent browser actions, page URL/title, and the shared relevance rubric. A candidate control is serialized into one per-control Noul question.
Gather visible controls, page context, task requirements, and recent actions.
When needed, ask Jev Noul questions about candidate controls and fit the most relevant set into the action request.
Ask Jev to choose among currently available operations and controls.
Code checks the action result and safety constraints before gathering the next observation.
The relevance pass is a bounded prefilter for browser controls, allowing the agent to spend its action request on more useful elements while keeping task policy and control execution in code.