Descriptions:
Sunil Pai, a long-time Cloudflare engineer and creator of the platform’s dynamic isolates, sits down at a developer conference for a technical conversation about Cloudflare’s emerging role in AI agent infrastructure. The discussion centers on two primitives Pai sees as foundational: Durable Objects, which he describes as the first actor-model implementation at the infrastructure layer — enabling millions of stateful serverless entities without spinning up full VMs — and Dynamic Workers, announced just weeks prior, which allow safe execution of LLM-generated code in sandboxed environments with zero startup time and configurable API exposure.
Pai uses Cloudflare’s own MCP server as a concrete example of this architecture. Rather than exposing all 2,600 of Cloudflare’s API endpoints as individual tools, the implementation provides just two tool calls — search and execute — where the LLM submits JavaScript code that runs in an isolate, enabling complex multi-step operations in a single round trip. He argues this approach is fundamentally different from patching existing infrastructure and positions Cloudflare to compete directly with Anthropic’s recently launched Claude Managed Agents platform.
The conversation also touches on the “slop fork” incident with Vercel — in which Pai used Claude Opus to generate a Cloudflare-compatible port of a Vercel Labs project over lunch, triggering an unexpected Twitter controversy — as an illustration of how AI-generated code is changing open-source collaboration norms. Throughout, Pai’s recurring theme is that the architecture of agentic software is still being discovered in real time, and that getting the underlying primitives right matters far more than shipping features on top of the wrong foundation.
📺 Source: Latent Space · Published May 24, 2026
🏷️ Format: Interview
