Cloudflare Targets Faster AI Agents with Dynamic Workers

Cloudflare is pushing a new approach to running AI agents at scale with the open beta of Dynamic Workers, a lightweight, isolate-based sandboxing system the company says is dramatically faster and more efficient than traditional containers.

According to Cloudflare, Dynamic Workers can start in milliseconds, consume only a few megabytes of memory, and even run on the same machine and potentially the same thread as the request that spawns them. When compared with Linux containers, the company claims this design makes Dynamic Workers roughly 100x faster to start and between 10x and 100x more memory efficient.

The launch builds on Cloudflare’s recent push around what it calls “Code Mode” for large language models (LLMs). Instead of chaining together one tool call after another, Cloudflare argues that LLMs can perform better when they are handed an API surface and asked to write code against it directly.

In that model, AI agents dynamically generate small code snippets to retrieve data, transform files, call services or automate workflows. Cloudflare says that converting a Model Context Protocol (MCP) server into a TypeScript API can reduce token usage by 81%, underscoring the company’s view that code-centric interactions can be more efficient than pure prompt-and-tool-call patterns.

Dynamic Workers are being positioned as the secure execution layer that makes this approach practical at scale. If AI agents are constantly writing and executing short-lived code fragments, the overhead of spinning up full containers or micro virtual machines quickly becomes a bottleneck. Cloudflare’s pitch is that isolates the underlying sandboxing mechanism behind Dynamic Workers are better suited to that high-churn, high-concurrency environment.

For enterprise technology leaders, Cloudflare frames this as more than a performance tweak. The company is effectively arguing that sandboxing itself is becoming a strategic layer in the AI stack, especially as organizations move toward architectures where many users may each have one or more agents continuously generating and executing code.

In that context, the economics and safety of the runtime start to matter nearly as much as the raw capabilities of the underlying model. If each agent interaction requires a fresh execution environment, every millisecond of startup time and every megabyte of memory matters. Cloudflare maintains that while containers and microVMs still have a role, they remain too heavy for this emerging pattern of fine-grained, per-request code execution.

The company situates Dynamic Workers within a longer evolution of secure code execution, noting that modern sandboxing has progressed through several generations of isolated runtime environments. Each generation has tried to build a “smaller, faster and more specialized” digital box for untrusted code. Dynamic Workers continue that trajectory by using isolate-based sandboxes designed to be compact enough to run alongside the request that created them, rather than as separate, heavyweight infrastructure.

By tying together its Code Mode concept, the token savings from exposing APIs as TypeScript, and the efficiency claims around Dynamic Workers, Cloudflare is making a clear case that if AI agents are going to write and run code as part of everyday workflows, the runtime must be as lightweight and scalable as the models themselves.