Forward Deployment Solves Specification Problem
David Mlcoch, co-founder & CEO of Asteroid, on browser automation and the last mile problem of AI
The hard part in enterprise browser agents is not getting a model to click buttons, it is turning messy human work into a precise set of instructions the model can repeat. In Asteroid's insurance and healthcare deployments, forward deployment engineers sit between the customer and the model, watch recordings, read SOP PDFs, test runs in Slack, and keep refining prompts and logic until the agent handles real edge cases like branching forms and missing fields.
-
Insurance quoting shows why this matters. A broker may answer 150 question forms where one answer changes the next screen, and the right input often depends on judgment learned on the job, not just the customer data. That judgment has to be translated into the agent's playbook.
-
This is the same pattern seen across enterprise AI agents more broadly. Companies like Sierra and Decagon use high touch deployment teams because adoption depends on custom integrations, workflow design, testing, and fast iteration from customer feedback to production, not just model access.
-
It also explains Asteroid's positioning versus developer tools like Browserbase, Stagehand, and Playwright. Those tools help engineers build automations, while Asteroid is trying to package the missing operational layer so non technical teams can stand up reliable agent workers for repetitive back office tasks.
As browser agents spread, more of the value will move from the raw model into the deployment muscle that captures workflow knowledge and turns it into reusable templates. The winners will be the companies that can convert one messy customer process into a repeatable playbook for an entire vertical, especially in insurance and healthcare.