Computer-Use Models Will Agentify the Web, Not APIs

The Fallacy of the API-First Web

The prevailing argument that AI agents will drive web interaction via APIs ignores the reality of the long tail of the web. With over 200 million active websites, most infrastructure is legacy, slow-moving, and designed for human consumption. Expecting these institutions to build and maintain MCP servers or standardized APIs is unrealistic. Many websites do not store state as text; they calculate and render it dynamically, meaning the source of truth is the visual output, not the underlying HTML or a hidden JSON endpoint.

Navigator: A Vision-Based Approach

The Navigator model demonstrates that the most general solution for web automation is to treat the browser as a rendering engine. By taking screenshots as input and outputting clicks, scrolls, and injected JavaScript, the model bypasses the need for site-specific API scaffolding. Navigator 1.5 achieves 97% accuracy on the Mind2Web benchmark, missing only 8 out of 300 trajectories. The model uses a hybrid approach: it clicks buttons like a human for navigation but executes custom JavaScript when that is more efficient, using the visual feedback loop as a formal verification system to confirm the task was completed successfully.

Scaling and Future Outlook

While critics point to latency and cost, these factors are rapidly improving through smaller, specialized model footprints. Current benchmarks show costs dropping to approximately 80 cents per task for complex 20-30 step interactions. The future of web "API-fication" will not be a formal protocol transition, but rather a layer of abstraction where orchestrators launch parallel browser instances to perform tasks, eventually presenting the result to the user in a structured format that functions as an API for all practical purposes.