Welcome to Agents Week

At Cloudflare, our mission has always been to help build a better Internet. This week, we are excited to kick off Agents Week, a dedicated event focused on preparing for the future of the Internet, particularly in the context of artificial intelligence (AI). The Internet was not originally designed for the age of AI, nor was the cloud infrastructure that supports it.

The Evolution of the Internet and Cloud Infrastructure

The cloud, as we understand it today, emerged from the last major technological shift brought about by smartphones. When smartphones made the Internet accessible to everyone, they didn’t just increase the number of users; they fundamentally changed the nature of online interactions. Users became accustomed to being always connected and expecting instant responses. This shift necessitated that applications handle significantly more users, leading to an evolution in the underlying infrastructure.

The industry responded with a straightforward approach: as the number of users increased, so did the number of application instances. As applications grew more complex, teams began to break them into smaller components known as microservices. Each team could then manage its own segment of the application, yet the core principle remained the same: a finite number of applications serving a multitude of users. In this model, scaling meant creating more copies of an application.

The Rise of Kubernetes and Containers

Kubernetes and containers became the standard tools for managing these applications. They simplified the process of spinning up instances, balancing loads, and disposing of unnecessary resources. Under this one-to-many model, a single instance could serve countless users, allowing for growth into the billions without overwhelming the management of resources.

Introducing Agents: A New Paradigm

However, the emergence of agents introduces a fundamental shift in this paradigm. Unlike previous applications, agents operate on a one-to-one basis. Each agent is a unique instance, serving a single user and executing one specific task. While traditional applications follow a fixed execution path, agents require their own execution environment, where a large language model (LLM) dictates the code path, dynamically calls tools, adjusts its methods, and persists until the task is completed.

A Personal Chef vs. a Restaurant

To illustrate this difference, consider the analogy of a restaurant versus a personal chef. A restaurant operates with a set menu and a kitchen designed for high-volume production. This is akin to most applications today. In contrast, an agent functions like a personal chef who asks, “What do you want to eat?” This chef may need different ingredients, utensils, or techniques for each meal, necessitating a unique setup for each task.

The Growing Demand for Agents

Over the past year, we have seen agents gain popularity, particularly coding agents, as developers tend to be early adopters of new technologies. Currently, most coding agents work by spinning up a container that provides the LLM with necessary resources such as a filesystem, git, bash, and the ability to execute arbitrary binaries. However, coding agents are just the beginning.

Expanding Accessibility

Tools like Claude Cowork are already making agents more accessible to non-technical users. As agents transition from being primarily used by developers to being utilized by a broader range of professionals—such as administrative assistants, research analysts, customer service representatives, and personal planners—the scale of demand becomes staggering.

The Mathematics of Scaling Agents

Consider the implications: if over 100 million knowledge workers in the United States each utilized an agent with approximately 15% concurrency, we would need capacity for around 24 million simultaneous sessions. Given that each CPU can handle between 25 to 50 users, this translates to a requirement of approximately 500,000 to 1 million server CPUs—just for the U.S. market, with one agent per person.

Now, envision each individual running multiple agents in parallel, and extend this scenario globally, where there are over 1 billion knowledge workers. The gap in computational resources needed is substantial.

Building Infrastructure for Agents

Eight years ago, we launched Workers, marking the beginning of our developer platform and a commitment to containerless, serverless computing. The initial motivation was practical: we required lightweight compute capabilities without cold starts for our customers who relied on Cloudflare for speed. Built on V8 isolates rather than containers, Workers turned out to be significantly more efficient—faster to start, cheaper to run, and inherently suited for the “spin up, execute, tear down” pattern.

What we did not anticipate was how well this model would align with the needs of the age of agents. While containers provide each agent with a fully equipped kitchen, isolates offer just the necessary tools for each specific task. They can be provisioned in milliseconds and cleaned up immediately after use. In a world where we need to support billions of ephemeral, single-purpose execution environments, isolates are the ideal solution.

Dynamic Workers and Efficiency

Recently, we took this concept further with the introduction of Dynamic Workers in open beta. These execution environments can be spun up on demand, taking just milliseconds to start and using only a few megabytes of memory. This efficiency allows for starting new instances for every request, running snippets of code, and discarding them—all at a scale of millions per second.

The Future of Agents

For agents to become widely adopted, they must also be affordable. Currently, running each agent in its own container can be prohibitively expensive, limiting agent tools primarily to coding assistants for engineers who can justify the costs. By utilizing isolates, we can run many more agents efficiently, making the economics viable at the scale required.

Transitioning to a New Era

While it is crucial to lay the groundwork for the future, we are still in the early stages of this transition. Every paradigm shift has a period where new technologies are forced to adapt to old models. The first cars were referred to as “horseless carriages,” and early websites were simply digital brochures. We are currently experiencing a similar phase with agents.

Bridging Two Eras

The Internet often exists in a state of flux between two eras. For example, IPv6 is superior to IPv4, yet the latter remains in use to avoid breaking existing systems. Cloudflare has always been dedicated to bridging these transitions, and the shift to agents is no exception.

Conclusion

As we embark on this exciting journey during Agents Week, we acknowledge the challenges and opportunities that lie ahead. The future of the Internet is being shaped by the rise of