🧭 Fara1.5-4B — Live Multi-Step Browser Agent
Fara1.5-4B by Microsoft Research AI Frontiers is a 4B vision-language computer use agent. This demo runs a real multi-step agent loop inspired by the Fara harness.
How it works — each step: ① screenshot the live page from a synced server-side browser (Playwright) → ② ask Fara for the next grounded action (click / type / scroll / navigate in a normalized 0–1000 space) → ③ map the coordinate onto a 1440×900 view and draw the marker → ④ apply it to the browser and re-render.
The run streams live: as each step completes, the big panel on the right updates with the current annotated frame and the log below fills in with Fara's reasoning + the exact action. When the loop finishes, the panel switches to an autoplaying slideshow of the whole sequence.
Step-by-step reasoning and actions will stream here as the agent runs.