APIs First Browser Agents Fallback
David Mlcoch, co-founder & CEO of Asteroid, on browser automation and the last mile problem of AI
This reveals that browser agents are not the new default interface for software, they are the workaround for software that never exposed clean machine access. An API lets an agent send a structured request and get back structured data in one step. A browser agent has to load pages, wait for elements, click through forms, and recover when a popup or layout change breaks the flow. MCP helps package either path for an LLM, but it does not remove the underlying cost and fragility of browser work.
-
In practice, APIs win on speed and reliability. OpenAI’s computer use docs still recommend human oversight and note meaningful reliability limits, which is why browser control fits short, repetitive tasks or legacy gaps better than core system integration.
-
The real demand for Asteroid comes from industries like insurance and healthcare, where workers still log into old portals, answer branching forms, and copy data between systems that have no usable API. That is the last mile problem, not a replacement for modern integrations.
-
The ecosystem is splitting by buyer. Browserbase sells hosted browser infrastructure and developer tools for engineers. Asteroid is pushing the layer above that, where an operations team can record a process, turn it into a reusable workflow, and call it through MCP when no API exists.
Over time, more software will expose agent friendly APIs and MCP servers first, because structured access is cheaper to run and easier to govern. Browser agents will remain essential in the large legacy surface area that cannot be rebuilt quickly, then gradually shrink into a fallback layer for the messy edges of enterprise workflows.