Lightpanda saves an agent’s browser work as a script that needs no model to run
Lightpanda’s browser agent can export a session as JavaScript, then replay it without an AI model. Its published speed figures are the maker’s benchmark, rather than a guarantee for your pages.
The agent figures out the browser workflow once. The next run can follow its saved script without calling a model. Lightpanda’s README describes that handoff through PandaScript, a JavaScript format with native browser operations.
During an agent session, /save exports the work. The lightpanda run command replays it. The project calls those scripts “deterministic and token-free,” meaning the browser procedure no longer needs an LLM at runtime. The website it visits can still change, reject a login or return different data.
That distinction matters more than the headline benchmark. The maker reports a 100-page workload using 123 MB peak memory versus Headless Chrome’s 2 GB, with execution times of five versus 46 seconds. Those figures belong to that benchmark. They do not measure your authenticated app or establish that every browser API your workflow needs is supported.
The next useful failure is a changed page. How clearly the saved script reports that change will determine whether this becomes a dependable routine or another silent scraper to babysit.