Matt Carrie and Sunil Pai from Cloudflare's agents team explain why Durable Objects turned out to be the right compute unit for AI agents: addressable, persistent, hibernating, stateful, and fast enough that 15ms London latency puts you inside a single animation frame. The Agents SDK builds on this to give resumable streaming, multi tab sync, and background scheduling out of the box, without any distributed systems engineering in userland.

The bigger reveal is Dynamic Workers: take a string of LLM generated code, run it in a sandboxed isolate with no ambient access, and grant only the capabilities you explicitly allow. They frame it as reclaiming 30 years of avoided eval. The session ends with both speakers teasing their afternoon talks, one on collapsing 2,600 Cloudflare API endpoints into a thousand token MCP tool, and a coding agent harness built entirely on Workers that they are, by their own admission, already shipping.

Speaker info:
- https://x.com/threepointone
- https://x.com/mattzcarey

AI Engineer

Cloudflare engineers Sunil Pai and Matt Carrie discuss their approach to building AI agents using Durable Objects as a stateful serverless compute primitive. They introduce 'Dynamic Workers' (eval++), a sandboxed execution environment that lets you run LLM-generated or user-generated code at scale without a full VM. Key capabilities covered include resumable streaming, multi-tab sync, MCP server hosting, and a worker-bundler that pulls NPM dependencies. They also tease upcoming features including a Cloudflare-native coding agent (similar to Claude Code), Python support in dynamic workers, a new CMS called M- with a plugin system built on dynamic workers, and a Next.js runtime on Workers. An 'Agents Week' announcement event is teased for the following week.

Why Eval++ Is the Next Great Compute Primitive — Sunil Pai & Matt Carrie, Cloudflare