Cloudflare Agents

Cloudflare Agents is a powerful, integrated platform designed for developers to build and operate agentic AI. Unlike traditional generative AI that simply responds to prompts, agentic AI introduces autonomy, goal-oriented reasoning, and adaptive decision-making. These agents can actively plan, iterate, and interact with their environment and external tools to accomplish complex, multi-step objectives. The entire platform is built on Cloudflare's robust global network, offering a unified solution for the complete agent lifecycle.

The core philosophy of Cloudflare Agents is to provide a seamless, scalable, and cost-effective environment. It breaks down the process of building an agent into four logical steps: receiving user input (via email, chat, or voice), reasoning with an AI model (either hosted on Cloudflare's Workers AI or through an external provider via AI Gateway), guaranteeing execution with state and compute (using Durable Objects and Workflows), and taking action by interacting with APIs and tools (like Browser Rendering, Vectorize, and D1 databases).

How to use Cloudflare Agents

Developing with Cloudflare Agents follows a structured, four-part workflow, all managed within the Cloudflare ecosystem:

1. Get User Input: Begin by capturing user requests or data. The platform supports various input channels, including Email with Email Workers, real-time chat with WebSockets and Pages, or even voice commands through Cloudflare Calls.
2. Ask AI for Reasoning: The agent processes the input and determines the next steps by querying a Large Language Model (LLM). Developers can choose to run optimized LLMs directly on Cloudflare's serverless GPUs with Workers AI or securely connect to third-party models like those from OpenAI or Anthropic using the AI Gateway.
3. Guarantee Execution: To ensure tasks are completed reliably, even if they are long-running, Cloudflare Agents combines state management with compute. Durable Objects provide persistent state for the agent, while Workflows orchestrate the sequence of tasks, allowing for retries, branching logic, and re-evaluation based on new information.
4. Take Action with Tools: Finally, the agent performs actions to achieve its goal. This involves using 'tools' which are structured ways to call APIs, manipulate data, or interact with systems. This can include rendering web pages for data extraction, querying a vector database with Vectorize, or reading/writing to a D1 SQL database.

Developers use the Cloudflare Agents SDK (demonstrated with JavaScript/TypeScript) to define agent behaviors, schedule tasks, and manage state in a programmatic way.

Core Features of Cloudflare Agents

• Unified Agentic AI Platform: All necessary components—from input handling and LLM inference to state management and action execution—are integrated into a single platform.
• Serverless LLM Inference: Run popular open-source LLMs on Cloudflare's serverless GPUs with Workers AI, paying only for what you use.
• Durable State & Execution: Leverage Durable Objects for persistent state and Workflows for reliable, long-running task orchestration.
• Cost-Efficient Pricing Model: Pay only for CPU time (actual computation), not wall-clock time (idle/waiting time), drastically reducing costs for I/O-heavy agentic tasks.
• WebSocket Hibernation: Maintain long-running WebSocket connections for real-time communication without the cost of a constantly active server, as Durable Objects hibernate when idle.
• Integrated Tooling Ecosystem: Seamlessly connect agents to other Cloudflare products like Vectorize for semantic search, D1 for SQL data, and Browser Rendering for web scraping.
• AI Gateway: Provides a secure, reliable, and observable gateway to connect to third-party LLM providers, offering caching, rate limiting, and analytics.

Use Cases for Cloudflare Agents

The platform's capabilities enable a wide range of sophisticated applications:

• Automated DevOps & Code Assistants: Create agents that can read GitHub issues, write code to fix them, and automatically open pull requests, as shown in the platform's own examples.
• Intelligent Customer Support Automation: Build chatbots that go beyond answering questions to perform actions like processing refunds, updating user accounts, or escalating complex issues with full context.
• Autonomous Research & Reporting: Deploy agents that can browse the web, gather information from multiple sources, synthesize findings, and generate a structured report.
• Complex Business Workflow Automation: Automate internal processes like employee onboarding, expense report processing, or lead qualification by integrating with various internal APIs and databases.

Advantages of Cloudflare Agents

Cloudflare Agents offers significant advantages for developers building next-generation AI applications. Its primary benefit is economic efficiency; the pay-for-CPU-time model is perfectly suited for the spiky, unpredictable nature of AI inference and agentic workflows, which often involve waiting for external APIs. This, combined with WebSocket Hibernation, leads to substantial cost savings over traditional hyperscalers. Furthermore, the unified platform simplifies the developer experience, removing the need to stitch together disparate services for state, compute, and AI. This accelerates development and reduces architectural complexity, all while benefiting from the performance and security of Cloudflare's global network.

Pricing and Plans

Cloudflare Agents operates on a pay-as-you-go pricing model, which is deeply integrated with the pricing of its underlying components like Cloudflare Workers, Workers AI, Durable Objects, and Vectorize. There is no single plan for 'Agents'; instead, you pay for the resources you consume. The key principle is cost-efficiency, where users only pay for the resources they actively use. This model includes:

• Compute: Billed based on CPU time, not wall-clock time, which is highly advantageous for tasks involving I/O waits.
• AI Inference: Charged per 1,000 neurons when using Workers AI.
• State & Storage: Costs associated with Durable Objects storage, D1 database usage, and Vectorize indexes.

This usage-based approach means there are no upfront commitments or hardware to manage. Cloudflare's ecosystem typically includes a generous free tier for its components, making it possible to start building and experimenting at no cost. Costs scale predictably as application usage grows.