Guide
Agent loop
Objective
Draw snapshot → ask Jev → act → snapshot. Stop in code. Rebuild the action space every step.
Jev does not drive Chrome. browser-use/jev-ultrafast (MIT) loops: observation → indexed table [1] button… [2] combobox…; operations CLICK / TYPE_TEXT / SELECT / SCROLL_* / WAIT / DONE / BLOCKED, offering only what the step supports. One request asks operation and each target in parallel (speculative fan-out). A small LLM writes text only on TYPE_TEXT. Published run: Zürich→London on Google Flights in ~7.073 s, ~$0.0039 in model cost (~98.4% Jev). Creative AI News checked the price against 90,558 input tokens at $0.042/M.
DeepWiki’s reading of that repo’s model.py: choose(state, goal, history) builds the payload; it checks IDs exist, probabilities sum to 1, confidence is finite in [0, 1]. Stop in code: DONE, a step cap, or the result is visible. MindStudio’s audit case: the tool said save failed on permissions, the agent claimed success — Jev marked the task failed and ~93% that the success claim was unsupported. The authors note ordinary code could catch that one; messier traces are untested at that scale.
The unlock is not just cheaper evals, but a tighter feedback loop for building reliable agents.
Exercise
A mini search page. In one step ask op / target / text in parallel. text is a closed set — do not ask it to invent the query. Read the distributions before you would click.