Last Mile Steering for Browser Automation
David Mlcoch, co-founder & CEO of Asteroid, on browser automation and the last mile problem of AI
The hard part in browser automation is not making a browser click once, it is turning a messy human workflow into something that works the same way every day. In practice, the agent has to handle popups, branching forms, missing fields, login state, and small page changes that break brittle scripts. That is why teams often spend weeks refining prompts, fallback logic, and supervision before an automation is reliable enough for production work like insurance quoting or healthcare data entry.
-
Stagehand sits in the developer tooling layer. It helps engineers build LLM driven browser automations, much like Playwright improved traditional browser scripting, but the workflow still depends on an engineer translating a domain expert's process into prompts, scripts, and exception handling.
-
The steering problem comes from the gap between a human SOP and what the browser actually presents. A broker may know that if one answer is yes, another field should usually be no, even when the customer never said it explicitly. That tacit judgment has to be encoded through many test runs and corrections.
-
Asteroid's wedge is packaging that iteration loop for operations teams instead of just developers. It records runs, surfaces live views, adds supervision, and turns a successful AI run into a reusable Playwright style script so the next run can be much faster and more deterministic.
This market is moving from raw agent capability to workflow reliability. The winning products will not just browse, they will capture domain rules, generate reusable scripts, and give teams a way to monitor thousands of runs in parallel. That is how browser automation shifts from demoware to real back office labor replacement.