Browser Automation as AI Last Mile

The key shift is that enterprise software is moving toward agent operated workflows on top of the same old systems, not a full rewrite of those systems. MCP helps agents talk to tools in a standard way, but it only works when a real API or tool endpoint exists. In the large installed base of insurer portals, medical scheduling systems, and other old web apps, browser agents become the adapter layer that lets AI complete the job by clicking, typing, and navigating for the user.

• MCP does not replace browser automation, it packages it. Asteroid describes MCP as a readable control layer on top of either APIs or browser agents. Browserbase has since productized the same idea with an MCP server that exposes browser actions like navigate, click, fill, and extract as tools an LLM can call.
  1 sacra 3 modelcontextprotocol 4 browserbase
• The product split is becoming clearer. APIs are fastest and cheapest when they exist. Browser agents are the fallback for systems with no API, and often the only path for real work in healthcare, insurance, and supply chain where staff still rekey information into brittle portals. That is why browser automation sits in the last mile of vertical AI.
  1 sacra 2 sacra
• This abstraction layer also changes who can automate work. Browserbase is aimed at developers running cloud browsers and Stagehand. Asteroid is pushing the layer upward to operations teams, where a broker or clinic operator can supervise an agent that fills forms and then save the generated workflow for repeated use, cutting a multi minute task down to seconds.
  1 sacra 5 anthropic 6 sacra

The next phase is a mixed stack. New software will expose clean APIs and MCP endpoints from day one, while old software remains agent operated for years. The winners will be platforms that can route each step to the cheapest reliable path, API when available, browser agent when necessary, and eventually broader desktop and spreadsheet agents for everything around the browser workflow.