The field of Artificial Intelligence is entering a new phase. After the rise of chatbots and conversational assistants, the focus is beginning to shift toward agents capable of reading context, reasoning, executing code, connecting to tools, and completing multi-step tasks more autonomously. In this scenario, Cloudflare has decided to strengthen its Agent Cloud offering with new infrastructure, security, and development components designed to take these agents from local demos to production workloads deployed across its global network. The company announced this on April 13 as an expansion of its platform, not as a single isolated feature.
The interesting part of the announcement isn’t just the commercial names, but the underlying ambition. Cloudflare argues that traditional infrastructure, based on containers or always-on virtual servers, doesn’t scale well in a world where each user or employee might have multiple agents running simultaneously. Its response involves combining ephemeral, cost-effective environments for executing AI-generated code, persistent spaces when an agent needs a “real” computer, storage designed for the code those agents generate, and a security layer integrated with its own network. In other words, Cloudflare aims to be the platform where models are invoked, but also where agents live, work, and connect to data and services.
From chatbots to agents with their own infrastructure
Cloudflare isn’t late to this conversation, but it’s elevating it to the next level. By 2025, it had already launched its Agents SDK for building persistent agents on Workers and Durable Objects, which already pointed to a clear thesis: the problem of agents isn’t just the model but also the environment where they run, their state, tools, and connectivity. With the expansion of Agent Cloud, the company takes a further step by organizing several pieces that were previously more scattered across its platform.
The most notable piece in terms of scale is Dynamic Workers, an isolated runtime designed to run AI-generated code on demand. Cloudflare explains that these environments start up in milliseconds, consume just a few megabytes of memory, and are approximately 100 times faster and up to 100 times more memory-efficient than a conventional container for these tasks. The idea is straightforward: if an agent needs to make an API call, transform data, or chain tools, it’s often unnecessary to spin up a full machine; a brief, secure, and cheap isolate suffices. For Cloudflare, this represents the key to scalable economics for large-scale agents.
This approach has significant implications for the AI market. Over the past two years, many agent-based tools have performed well in labs or premium products where per-user costs could absorb heavy infrastructure. But if agents are to become a common layer in enterprise software, the per-execution cost must decrease considerably. Cloudflare aims to address this bottleneck with an architecture more akin to extreme serverless than traditional container models per agent. This difference could matter far more than superficial interface improvements.
When agents need their own computer
Not everything can be handled with ephemeral environments. Some agents require cloning repositories, installing dependencies, running builds, maintaining background processes, or iterating over persistent filesystems. That’s where Cloudflare Sandboxes come in, now generally available. According to the official documentation, each sandbox functions as a persistent, isolated Linux environment with a shell, filesystem, background processes, a real terminal, preview URLs, snapshots, and a secure injection system for credentials so agents can access private services without directly receiving keys.
This is especially relevant for development agents and the so-called “code mode” systems. Cloudflare describes them as true computer spaces for agents: environments where a model can act more like an engineer than a chatbot, executing real loops of editing, testing, rerunning, and reviewing results. This approach is significant because it marks a clearer division in the market: one thing is serving models, and quite another is providing the infrastructure where those models can work persistently with operational context.
Complementing this layer is Artifacts, a new storage primitive compatible with Git, designed for the agent era. Cloudflare states it will enable the creation of tens of millions of repositories, fork from remote sources, and give agents a permanent home for code and data accessible via standard Git clients. It’s an intriguing idea because it points to a growing reality: as agents generate more code, branches, experiments, and artifacts automatically, storage and versioning will also need to adapt to this pattern.
Security, models, and the goal of becoming a full platform
The other critical aspect of the announcement is security. Cloudflare emphasizes that running agents quickly isn’t enough; they must also be isolated, secured with network control, and given safer access to data and private APIs. During the same Agents Week, the company also introduced Mesh, a way to provide private, segmented access for users, nodes, Workers, and agents via Workers VPC, and reinforced its messaging around authentication and internal access for agents. All of this aligns with a clear thesis: the future of enterprise agents will depend not only on models but also on combinations of runtime, identity, private networks, and tool control.
Furthermore, Cloudflare seeks to mitigate a growing risk: dependency on a single model provider. The company recalls that, after officially integrating Replicate in December 2025, it is expanding its catalog to combine in-house state-of-the-art models with open models under one interface. In its announcement, Cloudflare explicitly mentions OpenAI and includes a quote from its product team about deploying production-ready agents, emphasizing that switching providers should be as simple as changing a line of code. This is an ambitious promise but aligns with the current market pace, where models evolve too rapidly to commit indefinitely to one provider.
The core message is quite clear: Cloudflare doesn’t want to be just a network or security layer around AI applications. It aims to be the infrastructure where agents run, connect to private resources, store their code, and scale to millions of executions without the costs or complexity of traditional infrastructure. Whether it achieves this depends on real adoption, the maturity of these components, and how the market responds. Still, the move signals where the industry is headed: from isolated models to comprehensive platforms for agents.
Frequently Asked Questions
What is Cloudflare’s Agent Cloud and what is it for?
It’s Cloudflare’s solution to create, deploy, and scale AI agents across its global network, combining runtime, storage, security, and production-ready development tools.
What’s the difference between Dynamic Workers and Sandboxes at Cloudflare?
Dynamic Workers are designed for ephemeral, fast, and cost-effective execution of AI-generated code in isolates. Sandboxes, on the other hand, provide a persistent Linux environment with a shell, filesystem, and background processes when a full computer is needed.
What does Artifacts contribute to AI agents?
Cloudflare presents it as a Git-compatible storage solution designed for the agent era, capable of hosting code and data automatically generated at scale and integrating with standard Git clients.
Why is the integration of Replicate important for Cloudflare?
Because it allows expanding their model catalog and strengthens their strategy for a unified platform with less reliance on a single AI provider. Cloudflare confirmed that Replicate became part of the company on December 1, 2025.
via: cloudflare