Loading video...
Video Failed to Load
Introducing: Browser Use agents support WebMCP 💯 > UI is made for humans > Direct HTTP calls are made for agents > WebMCP makes interaction 100% deterministic No more flaky UI. Let's build the new web for web agents.
57,514 views • 1 month ago •via X (Twitter)
14 Comments

Moving to direct HTTP calls avoids flaky UI, but how does Browser Use handle sites without WebMCP support? Does it fall back to UI navigation?

I mean yeah it's just a helper that llm can use

How reliably does the LLM switch to UI navigation when that helper is not available on a site?

deterministic tool calls over flaky UI clicks is the only way browser agents stop being a demo. WebMCP feels like the missing layer.

Great to see it ! 💪

what about bot detection

Deterministic is the key word. UI automation always has that flaky layer, so a protocol that removes it could be a big shift for agent reliability. Curious how it handles auth or dynamic elements that still need a browser.

This framing is excellent. Humans get the UI. Agents get deterministic HTTP. That split alone removes a whole class of flaky "click the button" failure. Keeping a graceful human fallback beside the machine path is how these systems stay trustworthy when the happy path breaks.

Deterministic is the key word. UI automation always breaks on layout shifts, but a protocol layer gives agents a stable contract. Curious how you handle auth and state across sessions—that's where most agent workflows still get flaky.

Deterministic agent interfaces are the real unlock. Probabilistic clicking was always a demo. The harder question stays the same: once execution is reliable, who owns judgment when the agent can act faster than you can supervise?

hit this exact wall last month 3hrs chasing a race condition that was just a spinner loading 80ms slower on bad connections. did the determinism actually hold across different network conditions or are there still edge cases?

the 100% deterministic line is the real pitch. once a site exposes fixed tools the agent stops screenshot-guessing. leftover is still pages with no contract that need a live render or a login.

Deterministic beats "the agent figured out the DOM." Once the site exposes a contract, the eval is a function call, not a screenshot diff. Keep UI fallback as the exception path — with a receipt for every click that had to guess.

@sarah_edo WebMCP will play a strong role in rendering canvas based or iframe heavy sites

