From sandboxed code execution to Git-compatible storage, the company is positioning itself as the default infrastructure layer for a world where every user runs dozens of autonomous agents.
Cloudflare is expanding its Agent Cloud platform with a suite of new tools designed to help developers build, deploy, and scale AI agents at production grade — addressing what the company describes as a fundamental mismatch between the ambitions of agentic AI and the infrastructure currently available to support it.
The San Francisco-based company argues that existing approaches — relying on expensive, always-on virtual servers or isolated sandboxes — cannot scale to a world where every user and employee runs dozens of personal agents simultaneously. Its answer is a set of purpose-built primitives that trade heavy containerization for faster, cheaper, and more flexible alternatives.
“The way people build software is fundamentally changing. We are entering a world where agents are the ones writing and executing code,” said Matthew Prince, co-founder and CEO of Cloudflare. “Agents need a home that is secure by default, scales to millions instantly, and persists across long-running tasks.”
The centerpiece of the announcement is Dynamic Workers, an isolate-based runtime designed to run AI-generated code in a secure, sandboxed environment. Rather than spinning up a full container for each agent task, Dynamic Workers launch in milliseconds, execute the required JavaScript, and shut down — at what Cloudflare claims is 100 times the speed and a fraction of the cost of container-based alternatives, scaling to millions of concurrent executions without warm-up delays.
For storage, Cloudflare is introducing Artifacts, a Git-compatible primitive built for high-volume, agent-driven workloads. The tool allows developers to create tens of millions of repositories, fork from any remote source, and give agents a persistent home for code and data accessible to any standard Git client — addressing what the company says are growing scale and uptime limitations in traditional version control platforms.
Cloudflare is also announcing the general availability of Sandboxes, persistent Linux environments with a shell, filesystem, and background processes, for agents that require a full operating system. The environments allow an agent to clone a repository, install packages, run builds, and iterate in a tight feedback loop comparable to that of a human developer.
Rounding out the release is Think, a framework within the company’s Agents SDK designed for long-running, multi-step tasks. Current agents are often constrained by short lifespans that are poorly suited to complex, cross-platform work; Think is intended to address that limitation by giving developers tools for building agents with genuine persistence.
Also Read: When the CEO’s Avatar Speaks, Who Is Actually Responsible?
On the model side, following its acquisition of Replicate, Cloudflare is expanding its model catalog to include proprietary models from OpenAI alongside open-source alternatives, all accessible through a single interface. The company says switching between model providers requires changing only a single line of code.
OpenAI’s Rohan Varma, who leads product for Codex, said the partnership makes it significantly easier for developers to deploy production-ready agents powered by GPT-5.4 and Codex for enterprise workloads at scale.
