How do Cloudflare Agents differ from other agent frameworks?

Instead of building another library, Cloudflare focuses on providing the underlying infrastructure to run agents at scale. Their approach emphasizes using Durable Objects, which are uniquely suited for managing agent state and computation, aiming to simplify the complexity of agent development and deployment.

What is the core technology behind Cloudflare Agents?

The core technology is Cloudflare's Durable Objects, which are the world's first implementation of the actor model in infrastructure. This provides a robust and scalable foundation for running millions of agents, with compute and state closely tied together.

Can Cloudflare Agents handle scheduled tasks and background operations?

Yes, Cloudflare Agents support scheduling tasks, from simple reminders to complex cron patterns. These tasks can run in the background, even if the user or agent is offline, thanks to the persistent nature of Durable Objects.

What are the challenges in building and trusting AI agents?

A major challenge is the lack of a simple, trustworthy 'hello world' for agents. Unlike LLMs with RAG, agents often require sensitive credentials for complex tasks like making reservations, leading to a reliance on human-in-the-loop confirmation until trust is established.

How does Cloudflare's approach address the 'local first' movement?

Cloudflare's Durable Objects bring compute and data close to the user, potentially reducing the need for a separate 'local first' sync engine. They offer a 'best of both worlds' scenario where agents can maintain state and sync data continuously, even when the user is offline.

What are the future areas of focus for Cloudflare Agents?

Key areas include observability for managing large numbers of agents and making the technology accessible to non-technical users ('normies') for business automation. They are also exploring multimodal capabilities and richer integrations, like email and audio/video.

What new APIs has Cloudflare recently launched relevant to agents?

Cloudflare recently launched a direct REST API for browser rendering, allowing agents to interact with browsers similarly to how humans do, potentially replacing tools like Puppeteer. They are also working on Durable Object-enabled containers for more complex workloads.

Key Moments

npm install Agents — with Sunil Pai and Rita Kozlov (VP AI) of Cloudflare

Latent Space Podcast

Science & Technology5 min read36 min video

Mar 19, 2025|1,123 views|28|3

Save to Pod

Key Moments

TL;DR

Cloudflare launches Agents, leveraging Durable Objects for scalable AI agent infrastructure with a focus on developer experience and future multimodal capabilities.

Key Insights

Cloudflare Agents utilize Durable Objects, an actor model implementation, as a scalable container for running AI agents.

The platform emphasizes a developer-first approach, providing starter kits and documentation to facilitate agent development.

Key features include persistent state, background task execution, and scheduling capabilities, even when users are offline.

Cloudflare aims to bridge the gap in the current agent ecosystem by offering robust infrastructure rather than another framework.

Future development includes enhanced multimodal support (audio, video, email) and addressing challenges like observability and agent-to-agent communication.

The company is exploring solutions for authorization and authentication for agents acting on behalf of users, a significant challenge for web security.

INTRODUCTION TO CLOUDFLARE AGENTS

This discussion introduces Cloudflare's new Agents product, a significant launch leveraging their existing Durable Objects infrastructure. Rita Kozlov and Sunil Pai, from Cloudflare's VP AI and developer products respectively, explain that Agents are designed to provide a scalable platform for building and running AI agents. The core idea is to offer a robust infrastructure solution that simplifies the complex task of coordinating AI models, workflows, and external services, differentiating itself from the proliferation of agent frameworks in the market.

DURABLE OBJECTS AS THE FOUNDATION

A key enabler for Cloudflare Agents is Durable Objects, described as the world's first implementation of the actor model in infrastructure. These are lightweight compute instances that scale horizontally and can maintain state independently. This makes them ideal for hosting agents, providing persistent storage and reliable execution environments. The unique nature of Durable Objects allows agents to operate continuously, even if the user's device is offline, ensuring tasks are completed without interruption.

DEVELOPER EXPERIENCE AND GETTING STARTED

Cloudflare is prioritizing a strong developer experience with Agents. They provide publicly accessible code repositories, comprehensive documentation, and importantly, a starter kit. This starter kit includes a one-liner command to quickly spin up a functional agent with a frontend interface, demonstrating the ease of integration with existing JavaScript and React ecosystems. The emphasis is on enabling developers to rapidly prototype and deploy agents, reducing the typical friction associated with setting up complex development environments.

CORE FUNCTIONALITY AND FEATURES

The Agents platform offers several compelling features for agent development. It supports tool calls, allowing agents to interact with external services, and provides options for streaming text output. A notable inclusion is the 'human in the loop' confirmation step, which addresses trust and safety concerns by requiring user approval for certain actions. Furthermore, Agents incorporate scheduling capabilities, enabling tasks to be executed at specific times or intervals, even for users who are not actively connected.

ADDRESSING THE AGENT ECOSYSTEM CHALLENGES

Cloudflare acknowledges the current crowded agent landscape. Their approach is to focus on providing the underlying infrastructure and primitives, rather than another competing framework. A significant problem they aim to solve is the lack of a clear, economical 'hello world' for agents, contrasting with the widespread adoption of RAG for LLMs. By making Durable Objects more accessible for agent development, they hope to demystify the actor model and simplify the creation of agents that can reliably perform more complex, real-world tasks.

FUTURE DEVELOPMENT AND MULTIMODALITY

Looking ahead, Cloudflare is investing heavily in expanding the capabilities of Agents, particularly in multimodal interactions. While current examples focus on text-based communication and tool usage, future planned features include seamless integration with audio, video, and email. The team also highlights the importance of sophisticated agent-to-agent communication and observability tools to manage multiple agents effectively. The goal is to evolve Agents into a comprehensive platform that supports increasingly complex and human-like AI interactions.

NAVIGATING THE WEB SECURITY LANDSCAPE

A critical challenge discussed is how agents interact with web services, especially concerning authentication and authorization. As agents act on behalf of users, distinguishing them from malicious bots becomes paramount. Cloudflare is researching robust solutions for user identity and delegation, aiming to enable agents to perform actions like booking appointments while maintaining security. This is seen as a fundamental problem that needs solving for agents to gain widespread trust and adoption across various industries, from personal use to enterprise applications.

THE ROLE OF INFRASTRUCTURE IN APP DEVELOPMENT

The discussion touches upon the 'superficial hypothesis,' suggesting that available infrastructure heavily influences the types of applications developers can conceive and build. Cloudflare's strategy with Agents and Durable Objects is to provide powerful, accessible primitives that enable novel agentic workflows. By abstracting away complexities like Kubernetes and offering seamless scaling, they aim to make it easier for developers to build sophisticated applications that leverage AI agents, ultimately driving innovation and adoption across a broader range of use cases.

BROADER APPLICATIONS AND INDUSTRY IMPACT

The aspiration for Cloudflare Agents extends beyond typical developer use cases. The team is keen on seeing agents adopted by 'normies,' or everyday users and businesses, such as those in the hospitality industry. The vision includes agents automating tasks for small businesses, streamlining government services like DMV appointments, and integrating seamlessly into daily life. This focus on accessibility and real-world impact signifies Cloudflare's ambition to make AI agents a practical and widely utilized technology, not just a theoretical concept.

THE 'CHALLENGE ACCEPTED' FRONTIER

Cloudflare views significant challenges, like robust agent identification and handling heavy, CPU-intensive workloads, as opportunities. The hint at 'Durable Object enabled containers' suggests a future where complex operations and file systems can be managed within their infrastructure, potentially offering alternatives to Kubernetes. This forward-thinking approach aims to unify long-running processes with efficient resource utilization, making large-scale agentic workflows both feasible and cost-effective, further solidifying their commitment to solving complex infrastructure problems.

Mentioned in This Episode

●Products

●Software & Apps

●Companies

●Organizations

●Concepts

●People Referenced

Common Questions

Cloudflare Agents leverage Durable Objects, which are specialized compute instances, to provide a scalable platform for running AI agents. They allow developers to coordinate and communicate with agents using various protocols like HTTP, WebSockets, and email, without needing to manage their own code or infrastructure.

Topics

AI & Machine Learning Technology & Innovation Programming & Software AI Infrastructure Edge Computing Developer Tools Multi-Agent Systems Agent Frameworks

Mentioned in this video

Companies

Cloudflare

A web infrastructure and security company that has launched a new agents product. They are known for their developer-first culture and edge computing capabilities.

Anthropic

Mentioned in contrast to LangChain's work on multi-agent patterns, likely referring to Anthropic's research or products in a related area.

OpenAI

The underlying LLM provider used in the demonstration of the Cloudflare Agents starter kit, showing integration with major AI providers.

Browserbase

A company that likely provides browser automation tools for agents, mentioned in the context of how their efforts might be affected by Cloudflare's new browser rendering API.

People

Rita Kozlov

VP of AI at Cloudflare, involved in the launch of the new agents product. She has been with Cloudflare for nine years and leads product for developer offerings.

Sunil Pai

Lead product for developer offerings at Cloudflare, a key figure in the launch of the agents product. He has a history of building startups and was instrumental in developing the initial agent prototype.

Software & Apps

Durable Objects

A Cloudflare feature that implements the actor model for JavaScript developers, providing scalable and horizontally distributed compute instances. It serves as the core infrastructure for running agents.

LangChain

A framework mentioned as an alternative or existing toolset for building agents, contrasting with Cloudflare's approach of providing infrastructure rather than a competing library.

AISC SDK

The Software Development Kit used in the starter kit for the Cloudflare Agents, noted for being a good choice for the example.

Kubernetes

Container orchestration system, discussed as a complex infrastructure that Cloudflare aims to abstract away for agents, offering a simpler alternative.

Guardrails

A feature integrated into Cloudflare's AI Gateway, described as 'LL Guard', likely for enforcing security and compliance rules on LLM interactions.

Puppeteer

A Node.js library for controlling headless Chrome or Chromium, mentioned as a tool that developers previously used for browser automation before Cloudflare's new API.

React

A JavaScript library used in the starter kit for the front-end of the agent application.

Concepts

Actor Model

A conceptual model for concurrent computation, implemented by Cloudflare's Durable Objects. It's highlighted as being perfectly suited for running agents due to its ability to manage state and scale horizontally.

Products

AI Gateway

A Cloudflare AI product that recently integrated 'Guardrails', a feature for managing LLM behavior, indicating continued development in Cloudflare's AI offerings.

Organizations

DMV

Department of Motor Vehicles, used as a prime example of a government service that agents could one day manage, highlighting the potential for broad application.