One live browser session, a real signed-in Duolingo tab, a Hindi lesson running through E2LLM MCP. Claude Opus 4.8 starts. At 2:09 it is stopped and the task is handed to GPT-5.5 Medium; at 4:35, to Nemotron 3 Ultra via Perplexity. Each handoff is manual and on screen. Same tab, same browser state, no restart - the models share no context beyond the prompt and what they read from the page.
Each model receives the same structured SiFR view of the page it encounters: roles, text, actionable elements, selectors, positions. That representation is what lets the next model continue mid-task. The reasoning and the browser actions are side by side in the frame, including the Devanagari segment.
The recording runs uncut - the swaps, the prompts, and the moment the disk fills up and stops the recording are all on screen, with 12 correct answers in a row at that point.
SiFR spec (MIT): https://github.com/e2llm/e2llm-sifr
Setup guide: https://e2llm.com/docs/setup/
Top comments (0)