Asteroid's AI-to-Playwright automation
David Mlcoch, co-founder & CEO of Asteroid, on browser automation and the last mile problem of AI
This turns one successful AI run into a reusable automation asset. Instead of making the model rediscover every click and field on each run, Asteroid lets the agent do the task once, then hands back a parameterized Playwright script that can be rerun with new inputs. That is the bridge from slow, exploratory agent behavior to fast, repeatable production workflows, especially for teams filling the same insurance, healthcare, or back office forms over and over.
-
The key product move is exposing both layers, natural language for the first run, and code for repeat runs. A non technical operator can say fill this form with this data, while a technical team can inspect or edit the generated Playwright script and wire it into a larger workflow.
-
Playwright matters because it is the stable execution layer under the AI. It supports reliable browser actions with built in waiting and element checks, which makes a generated script much less fragile than a raw sequence of clicks. That is what makes turning an agent session into reusable code practical.
-
This also explains the speed jump from minutes to seconds. The first pass uses AI to explore the page, handle branches, and figure out where everything is. Later runs skip most of that discovery work and just inject new variables into the saved script. Browserbase describes a very similar pattern in its Director product.
The market is heading toward a split model, where AI handles discovery and exception handling, while generated scripts handle the bulk of repeated work. The winners will be platforms that can capture a successful human or agent workflow once, convert it into maintainable code, and run thousands of those jobs in parallel inside hosted browsers.