Michele Catasta, VP of AI at Replit, presents a bold vision of full autonomy for coding agents aimed at non-technical users. He introduces a third dimension to Swyx's well-known chart: beyond latency and autonomy, accessibility to non-developers must be considered.
Catasta distinguishes two types of autonomy using an automotive analogy. Supervised autonomy (Tesla FSD) requires a "driver's license" - the user must be technically competent to intervene. Full autonomy (Waymo), which Replit targets, allows users to "sit in the back seat" with no technical skill required. This vision aims to democratize software creation for all knowledge workers.
The "Painted Doors Problem" is a major challenge: 30% of features generated by agents are defective on the first attempt, creating non-functional dummy interfaces. Non-technical users can neither diagnose the issue nor provide the technical feedback needed. Replit's solution: autonomous testing, which lets agents verify their own work without human intervention.
The technical evolution unfolds across three generations of agents. The first, React-based, offered basic autonomy. The second integrated native tool calling. The third generation, recently launched (B3), breaks the one-hour autonomy barrier while remaining coherent on long-running tasks. This autonomy remains scoped: the agent makes all technical decisions while the user retains control over the final product.
The code verification spectrum has evolved from static analysis (LSP) toward more sophisticated approaches: execution, unit tests, API tests, up to browser testing. Replit implements a hybrid approach combining Computer Use (direct but costly interaction) and Browser Use (DOM manipulation, more efficient) to automatically test generated web applications.
The presentation reveals that autonomy does not necessarily mean slowness. An agent can be fast AND autonomous if the task scope is narrow. Users retain control over what they build, not how it is built. This separation of concerns allows knowledge workers to create software without technical expertise.
Catasta concludes that this transformation represents a fundamental shift: moving from coding assistance to the full delegation of technical implementation, opening software creation to millions of new potential creators.