What we kept seeing
Every framework spoke the wrong language.
Playwright, Selenium, Cypress — mature tools built for humans driving a browser. When AI agents arrived as a real development workflow, the natural move was to plug them in. The results were bad. Token payloads were bloated. IPC round-trips killed the agent loop. There was no concept of intent, no NLP understanding, and nothing stopping an agent from clicking Delete on a production record.
The architecture wasn't wrong for humans. It was the wrong architecture entirely for agents.