Descriptions:
Cloudflare engineers Sunil Pai and Matt Carrie take the stage at the AI Engineer conference to explain how Cloudflare’s Durable Objects infrastructure turned out to be a near-perfect compute primitive for AI agents. Unlike stateless serverless functions, Durable Objects maintain persistent state, run long-lived background tasks, hibernate when idle, and address all requests for a given entity to a single instance — properties that map directly onto the requirements of reliable agent workloads. The team has been building production agents on this foundation for over a year.
The centerpiece of the talk is Cloudflare’s new “dynamic workers” capability, internally dubbed eval++ — a secure, sandboxed runtime for generating and executing code at request time. The engineers argue that eval-style dynamic execution has been off-limits for 30 years due to safety concerns, but Cloudflare’s isolated worker environment reopens that branch of the tech tree cheaply and safely. In a world where every user can generate code with AI, this changes how interfaces and agent capabilities can be expressed.
The talk also previews a forthcoming deep-dive on “code mode” for MCP servers: rather than enumerating Cloudflare’s 2,600 API endpoints as static tool schemas, the approach generates executable code on demand, compressing the entire API surface to roughly 1,000 tokens. Both dynamic workers and code-mode MCP are presented as solutions to context bloat and tool-list explosion — two of the most common failure modes in production agent deployments.
📺 Source: AI Engineer · Published June 08, 2026
🏷️ Format: Deep Dive
