Welcome to Hyeonggeun's Terminal!
2026. 10. 26. 4:52:36 AM

← Back to projects
$>

2026–present · #product · #ongoing

Web Runtime for AI Agents

Herbert Computer

An execution layer for AI agents that turns completed web workflows into reusable REST and MCP APIs, rerun with new parameters without further AI reasoning.

About this project

I developed an automated pipeline that turns natural-language website tasks into reusable, parameterized APIs. The system guides coding agents through browser exploration, workflow implementation, and execution-based verification, then packages successful workflows into typed interfaces with structured inputs and outputs. It automatically checks input requirements, requests clarification when needed, and verifies reuse with alternative inputs before publishing the generated capabilities to a shared library.

In the best-performing internal evaluation configuration, the pipeline generated and execution-validated reusable APIs in approximately 7.5 minutes per task on average (minimum: 3.5 minutes; maximum: 12.5 minutes), including browser exploration, implementation, repairs, and verification with both original and alternative inputs. Across five website tasks with three repetitions each, 14 of 15 runs passed execution verification. The average gives each task equal weight using its successful-run mean; the minimum and maximum refer to individual successful runs. Once generated, these APIs can execute the workflow with new parameters without requiring an LLM to reason through each browser interaction again.

In a job-posting task with Claude Code, the runtime completed the task about five times faster using 78% fewer tokens than Playwright MCP. The first two videos show job posting with Playwright and Herbert Runtime, respectively.

I demonstrated the runtime through cross-site job posting and a TheVC-style startup database assembled from more than ten web sources.

The related AI-native Full Self Driving Browser focuses on the browser experience and agent execution of tasks across websites.

← Back to projects