agent-browser adds saved flows, browserless replay

Spinning up headless browser instances for repetitive agent queries is slow, fragile, and burns unnecessary context tokens. Saved flows bridge the gap between ad-hoc web exploration and reliable production automation by compiling recorded browser sessions into lightweight, deterministic requests.

• –**Tiered execution strategy**: Replays attempt direct HTTP requests first, degrade to page-level evaluation, and only fall back to CLI replay if necessary, automatically caching whichever tier succeeds.
• –**Context window efficiency**: Compressing multi-kilobyte pages into compact byte-sized payloads prevents context bloat when agents handle frequent web monitoring tasks.
• –**Parameterized re-execution**: Agents can supply dynamic input arguments on subsequent runs to check prices, order statuses, or metrics without re-navigating site layouts.
• –**Shift toward deterministic agent tooling**: Moving from continuous DOM navigation to a "record once, replay deterministically" paradigm offers a scalable pattern for production-grade web agents.