SUMM

Cloudflare serves about 20% of all internet traffic, with products touching end-users even if they haven’t directly signed up
In addition to traditional services like CDN and DNS, Cloudflare offers developer resources such as functions, storage, compute, and AI inference
The mission is to make it very easy for developers to bring ideas to life, from writing code to deploying for millions of users
A significant paradigm shift toward AI adoption is underway, with rapid recent growth
AI usage stats: a year ago, 44% of developers and projected 50% of knowledge workers by 2030 were using AI; current numbers already exceed 75% for both groups
The transition in AI workloads has moved from mainly training to a majority now focused on inference, reflecting current industry trends

The next phase after training and inference is automation via agents
Agents can go beyond simple AI augmentation (like generating an email) to handle complex workflows (gathering lists, drafting, approvals, monitoring replies)
Early business results include 20% revenue increases with sales automation agents, 90% faster support response times, and 50–75% time savings in applicable processes
Agents are already reshaping how people work and unlock new productivity gains

An agent consists of: (1) client interface for human interaction, (2) AI reasoning for logic, (3) workflows for keeping track of actions, and (4) access to tools (APIs, web, databases)
Practical example: a voice agent would use WebRTC, speech-to-text, an LLM for reasoning, workflow tracking, and tool integrations; may also require a human in the loop for approvals

Anthropics' MCP (Multi-Client Protocol) is a new standard for exposing APIs to LLMs using natural language, introducing efficient tool-calling methods
MCP servers support multiple clients, resources (files, DB records), prompts (for agent customization), tool integration, and (theoretically) sampling
MCP implementation challenges involve transport protocols, authentication, and memory management
Cloudflare’s agents SDK provides out-of-the-box features to simplify MCP server deployment: MCP agent class, built-in auth, HTTP streaming, real-time websocket communication, React integration, and chat capabilities
Cloudflare's durable objects enable stateful, serverless functions, making state management easy without traditional database setup

Deploying an MCP server on Cloudflare is quick: define an MCP class, connect tools, persist user preferences and state automatically
Example: building a personalized book recommendation agent that improves over time with persistent memory, regardless of client platform
Benefits include eliminating the need to manually manage databases, connections, or scaling—Cloudflare handles infrastructure and scaling

Agents benefit from workflow coordination, managing state across chains of tools and human intervention
Handling human-in-the-loop tasks means dealing with long-running asynchronous workflows (potentially minutes to months), requiring persistence and the ability to resume tasks
Use case: customer Knock implemented agent workflows for credit card approvals, integrating chat, deferred tool calls awaiting human approval, and stateful event routing, preventing duplicate actions through built-in state management

MCP servers, once deployed, can be accessed by multiple client interfaces—developers’ tools like Cursor, Claude, and ChatGPT support remote MCP servers natively
Developers can also build custom apps or voice interfaces that interact with MCP servers, with Cloudflare tools translating WebRTC to websockets for seamless integration

Building effective agents requires four key building blocks: client, AI model, workflow management, and tool integration
Cloudflare’s agents SDK offers a rapid and accessible way to build and deploy agents for a variety of workflows

Building Agents (the hard parts!) - Rita Kozlov, Cloudflare