2026-03-14 · Lab Notes ◆ Experimental

Browser Agents

Playwright MCP made browser automation a standard agent tool. Reliability is the remaining problem.

Success rate85–92%

Median task30–90s

5-step workflow59% compound

What MCP Changed

Before DOM scraping · brittle selectors · per-framework integration

→

After ARIA snapshots · typed actions · standardized MCP protocol

The agent sees semantic page structure, not raw HTML. It clicks based on what an element does, not where it sits.

Form filling

92%

85%

Data extraction

88%

SPA navigation

70%

Multi-step workflows

60%

Each step at ~90% reliability. A 5-step workflow compounds to 59%. This is the ceiling.

Timing SPA content loads after page event. Agent clicks too early.

Detection Cloudflare/Datadome flag automated browsers. Arms race.

Noise Cookie banners, modals, popups — agent can't dismiss reflexively.

Recovery No human intuition to reroute when an action fails.

Vision modelsScreenshot + click fallback

Agent memorySite-specific patterns

Session replayRecord → replay → explore

Browser automation is becoming a standard agent capability, not a product category. The terminal equivalent for the web.