Browser-Agent records nav, replays zero LLM calls

This is a practical attack on the most wasteful part of browser agents: re-paying tokens for the same navigation path over and over. The real value is not just token savings, but making browser automation feel deterministic enough to schedule, cache, and reuse.

• –Separating stable navigation from live extraction is the right abstraction; most browser-agent failures and costs come from redoing the same clicks, scrolls, and page transitions.
• –The `aiFallback` idea is the important safety valve: you preserve replay economics while still recovering from page drift without re-running the whole flow.
• –Backend portability matters here. Supporting local Chromium, AWS Lambda via `@sparticuz/chromium`, and remote CDP endpoints makes the library usable for dev, scheduled jobs, and scraping infrastructure.
• –This sits closer to infrastructure than a flashy end-user product: useful for teams building scraping, monitoring, QA, or data-extraction pipelines that need repeatable browser runs.
• –The limitation is obvious but acceptable: replay works best when the page structure is stable, so it reduces token spend rather than eliminating the need for AI entirely.