Cloudflare's Durable Objects and Dynamic Workers as AI agent primitives
Matt Carrie and Sunil Pai from Cloudflare's agents team explain how Durable Objects became their core compute unit for AI agents, and preview Dynamic Workers — sandboxed isolates that safely execute LLM-generated code.
Score breakdown
Dynamic Workers extend the Durable Objects model to safely execute LLM-generated code in isolated sandboxes, addressing one of the core trust and safety challenges in agentic systems that run arbitrary model output.
- 01Durable Objects are addressable, persistent, hibernating, and stateful serverless compute units that Cloudflare's agents team uses as the core primitive for AI agents.
- 02Cloudflare's London network delivers approximately 15ms latency — within a single 60 FPS animation frame (~16ms).
- 03The Agents SDK provides resumable streaming, multi-tab sync, and background scheduling without distributed systems engineering in userland.
Matt Carrie and Sunil Pai, who work on the agents team at Cloudflare, open by explaining the core problem with standard serverless functions: they are stateless, so any in-memory counter or session state is lost between requests. Durable Objects solve this by giving a single class instance a persistent identity — every request and WebSocket connection for a given ID lands in the same place. They highlight that Cloudflare's network delivers roughly 15ms latency in London, which they note is within a single animation frame at 60 FPS, enabling real-time collaborative experiences like TLDraw. The Agents SDK builds on Durable Objects to offer resumable streaming, multi-tab sync, and background scheduling out of the box, removing the need for userland distributed systems work.
The speakers frame this as reclaiming 30 years of programming culture that avoided `eval`.
The bigger reveal in the session is Dynamic Workers: the ability to take a string of LLM-generated code, run it inside a sandboxed isolate with no ambient access, and grant only the capabilities explicitly permitted by the caller. The speakers frame this as reclaiming 30 years of programming culture that avoided `eval`. The session ends with both speakers teasing their individual afternoon talks — one covering how Cloudflare collapsed its 2,600 API endpoints into a roughly thousand-token MCP tool, and another on a coding agent harness built entirely on Workers that they describe as already shipping.
Key facts
- 01Durable Objects are addressable, persistent, hibernating, and stateful serverless compute units that Cloudflare's agents team uses as the core primitive for AI agents.
- 02Cloudflare's London network delivers approximately 15ms latency — within a single 60 FPS animation frame (~16ms).
- 03The Agents SDK provides resumable streaming, multi-tab sync, and background scheduling without distributed systems engineering in userland.
- 04Dynamic Workers run LLM-generated code strings in sandboxed isolates with no ambient access; capabilities must be explicitly granted.
- 05The speakers frame Dynamic Workers as reclaiming 30 years of avoided `eval`.
- 06One upcoming talk covers collapsing Cloudflare's 2,600 API endpoints into a roughly thousand-token MCP tool.
- 07A coding agent harness built entirely on Workers is described as already shipping.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 9, 2026 · 17:05 UTC. How this works →