Cloudflare Agent Cloud 2026 — Building AI Agents at the Edge with Workers, Durable Objects, and Project Think

Posted on: 4/18/2026 2:09:47 AM

1. Cloudflare — From CDN to Agent Cloud

Cloudflare has been on a remarkable journey: from a pure CDN and web-security provider to a full-fledged platform for AI Agents. With Agents Week 2026 (April 2026), Cloudflare officially introduced the Agent Cloud concept — a vision that turns its global edge network into infrastructure for distributed, stateful, serverless AI agents.

The core differentiator: instead of renting a 24/7 VM to host each agent, Cloudflare lets agents hibernate when idle and only consume resources during actual work — pushing running costs for idle agents close to zero.

2. Workers & Dynamic Workers — V8 Isolates at the Edge

2.1. Cloudflare Workers: the serverless edge platform

Cloudflare Workers runs JavaScript/TypeScript on V8 isolates — the same engine powering Chrome — at more than 330 edge locations worldwide. Not containers, not VMs — each request is handled inside a lightweight isolate with near-zero cold start.

Key technical characteristics:

• Free tier: 100,000 requests/day, 10 ms CPU time per invocation
• Paid ($5/month): 10 million requests, 30 s CPU time, unlimited static assets
• Supported languages: JavaScript, TypeScript, Python, Rust (via WASM)
• Bindings: direct connections to KV, R2, D1, Queues, and Durable Objects with no network hop

2.2. Dynamic Workers: code-at-runtime

Dynamic Workers (open beta in March 2026) is a major step forward: it allows code to be injected and executed at runtime via an API, with no prior deployment. It's the foundation for AI-generated code — the agent writes code, pushes it, and the code runs instantly.

Dynamic Workers follow the zero-ambient-authority security principle:

// A Dynamic Worker starts with no permissions at all
const worker = await createDynamicWorker({
  code: agentGeneratedCode,
  bindings: {
    globalOutbound: null,  // No network access
    // Only grant what's needed via bindings
    DB: env.MY_D1_DATABASE,
    STORAGE: env.MY_R2_BUCKET,
  }
});

3. Durable Objects & Facets — State for AI Agents

3.1. Durable Objects: single-threaded actor model

Durable Objects solve serverless's biggest challenge: state. Each Durable Object is a single-threaded actor with:

• Its own SQLite database on local disk — near-zero latency
• Transactional storage for consistency
• Hibernation: it sleeps when idle and wakes on request — zero cost while inactive
• WebSocket support: maintains real-time connections

The economic implications are clear: if you have 10,000 AI agents but only 1% are active at once, a traditional VM setup needs 10,000 instances running continuously. With Durable Objects + hibernation you only need ~100 active instances at any given moment.

3.2. Durable Object Facets: isolation for dynamic code

Facets (Agents Week 2026) extend Durable Objects with a parent–child model:

graph TD
    A["🏗️ Parent Durable Object
(Platform code)"] --> B["📊 Parent SQLite
Metadata, billing, logs"]
    A --> C["🔒 Facet (Child)
AI-generated dynamic code"]
    C --> D["💾 Child SQLite
Application data"]
    B -.->|"❌ Isolated"| D
    A --> E["🔒 Facet (Child 2)
Another application"]
    E --> F["💾 Child SQLite 2
Separate data"]
    D -.->|"❌ Isolated"| F

    style A fill:#e94560,stroke:#fff,color:#fff
    style C fill:#2c3e50,stroke:#fff,color:#fff
    style E fill:#2c3e50,stroke:#fff,color:#fff
    style B fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style D fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style F fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    

Each Facet has its own SQLite database, fully isolated from the parent and from other Facets. The parent controls rate limiting, quotas, and billing — while the child focuses purely on application logic. This is the foundation for platforms that let AI create applications with persistent state.

export class AppRunner extends DurableObject {
  async fetch(request: Request): Promise<Response> {
    // Load dynamic code from R2 or an API
    const appCode = await this.env.R2.get("apps/user-123/code.js");

    // Create a facet — each app gets its own SQLite
    const facet = this.ctx.facets.get("user-app-123", {
      className: "UserApp",
      code: await appCode.text(),
    });

    return facet.fetch(request);
  }
}

4. Project Think — the Next-Generation Agents SDK

Project Think is Cloudflare's official framework for AI agents, built on top of Durable Objects. Instead of gluing primitives together yourself, Think provides a base class that handles the full lifecycle of an agent.

4.1. Core architecture

import { Think } from "@cloudflare/agents";
import { createWorkersAI } from "@cloudflare/agents/ai";

export class MyAgent extends Think<Env> {
  getModel() {
    return createWorkersAI({ binding: this.env.AI })(
      "@cf/moonshotai/kimi-k2.5"
    );
  }
}

4.2. Key primitives

Durable Execution with Fibers:

Fibers allow agent loops to run for many minutes (or longer) without losing progress. Each fiber is recorded in SQLite before it runs, can checkpoint at any time, and recovers automatically if the platform restarts.

await this.runFiber("research-task", async (fiber) => {
  const results = await this.searchWeb(query);
  await fiber.stash(); // Checkpoint — safe against crashes

  const analysis = await this.analyzeResults(results);
  await fiber.stash(); // Second checkpoint

  return this.generateReport(analysis);
});

Sub-agents via Facets:

Each sub-agent is a child Durable Object with its own SQLite, communicating over typed RPC. The parent agent delegates work to sub-agents — each running isolated and able to hibernate independently.

Persistent Sessions:

Conversations are stored as a tree (parent-message relationships), with non-destructive compaction (summarize rather than delete) and full-text search via SQLite FTS5. Sessions can be forked to explore multiple directions without losing the original context.

4.3. The Execution Ladder — 5 tiers

5. Workers AI & AI Gateway — Unified Model Access

5.1. Workers AI: 50+ models at the edge

Workers AI provides inference for 50+ open-source models directly on Cloudflare's GPU network. No infrastructure to manage — call a model via a binding just like calling a function:

const response = await env.AI.run(
  "@cf/meta/llama-4-scout-17b-16e-instruct",
  {
    messages: [
      { role: "user", content: "Analyze microservices architecture" }
    ]
  }
);

Notable recent models (April 2026):

• Google Gemma 4 26B A4B — MoE with 26B total / 4B active, 256K context, supports vision + thinking + function calling
• GLM-4.7-Flash — 131K-token context, optimized for summarization
• Qwen3-30B-A3B — MoE activating only 3B parameters per forward pass
• EmbeddingGemma-300M — 768-dim vectors, optimized for low-latency embedding

5.2. AI Gateway: a unified proxy for every AI provider

AI Gateway acts as a unified inference layer, supporting 14+ providers (OpenAI, Anthropic, Google, Mistral, …) through a single interface. The new breakthrough: the same AI.run() binding works for both Workers AI models and third-party models.

graph LR
    A["🤖 AI Agent"] --> B["🌐 AI Gateway"]
    B --> C["Workers AI
Llama, Gemma, Qwen"]
    B --> D["OpenAI
GPT-4.1, o4"]
    B --> E["Anthropic
Claude Opus, Sonnet"]
    B --> F["Google
Gemini 2.5"]
    B --> G["Caching Layer"]
    B --> H["Rate Limiting"]
    B --> I["AI Firewall"]

    style A fill:#e94560,stroke:#fff,color:#fff
    style B fill:#2c3e50,stroke:#fff,color:#fff
    style C fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style D fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style E fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style F fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style G fill:#f8f9fa,stroke:#4CAF50,color:#2c3e50
    style H fill:#f8f9fa,stroke:#ff9800,color:#2c3e50
    style I fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    

AI Gateway provides:

• Caching: cache responses for identical prompts, reducing cost and latency
• Rate limiting: control quotas per user/key/endpoint
• AI Firewall: detect prompt injection and data exfiltration before requests reach the model
• Analytics: dashboards for token usage, latency, and error rate per model/provider
• Fallback: automatically switch to another provider when one has issues

6. MCP Server at the Edge

Cloudflare has partnered closely with Anthropic to build infrastructure for remote MCP servers — bringing the Model Context Protocol to the edge with scale and hibernation.

An enterprise MCP architecture on Cloudflare:

graph TD
    A["AI Agent / Claude"] -->|"MCP Request"| B["Cloudflare Access
OAuth 2.0 / RFC 9728"]
    B --> C["AI Gateway
Code Mode reduces tokens"]
    C --> D["MCP Server Portal
(McpAgent on DO)"]
    D --> E["Tool: Database Query"]
    D --> F["Tool: File Storage"]
    D --> G["Tool: External API"]
    D -.->|"Hibernate when idle"| H["💤 State preserved
in SQLite"]

    style A fill:#e94560,stroke:#fff,color:#fff
    style B fill:#2c3e50,stroke:#fff,color:#fff
    style C fill:#2c3e50,stroke:#fff,color:#fff
    style D fill:#e94560,stroke:#fff,color:#fff
    style E fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style F fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style G fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style H fill:#f8f9fa,stroke:#4CAF50,color:#2c3e50
    

Standout features:

• Code Mode: reduces token cost by compressing tool descriptions
• Managed OAuth: implements RFC 9728 for agent authentication without building your own OAuth flow
• Cloudflare Mesh: grants MCP servers access to private databases/APIs without manual tunneling
• Scannable API tokens: resource-scoped permissions that honor least-privilege

7. The supporting service ecosystem

7.1. Storage Layer

7.2. Agents Week 2026 — New services

8. Reference architecture: an AI Agent on Cloudflare

Here's an end-to-end architecture for a production AI agent running entirely on Cloudflare:

graph TD
    subgraph "Client Layer"
        U["👤 User"] --> W["Worker
(API Gateway)"]
    end

    subgraph "Agent Layer"
        W --> T["Think Agent
(Durable Object)"]
        T --> F1["Sub-agent 1
(Facet - Research)"]
        T --> F2["Sub-agent 2
(Facet - Code Gen)"]
        T --> F3["Sub-agent 3
(Facet - Review)"]
    end

    subgraph "AI Layer"
        F1 --> AI["Workers AI
Llama 4 / Gemma 4"]
        F2 --> GW["AI Gateway
→ Claude / GPT"]
        F3 --> AI
    end

    subgraph "Tool Layer"
        T --> MCP["MCP Server
(McpAgent on DO)"]
        MCP --> BR["Browser Run"]
        MCP --> SB["Sandbox"]
        MCP --> EX["External APIs"]
    end

    subgraph "Storage Layer"
        T --> D1["D1
Agent metadata"]
        T --> R2["R2
Files & artifacts"]
        T --> VZ["Vectorize
RAG embeddings"]
        T --> Q["Queues
Async tasks"]
    end

    style U fill:#e94560,stroke:#fff,color:#fff
    style W fill:#2c3e50,stroke:#fff,color:#fff
    style T fill:#e94560,stroke:#fff,color:#fff
    style F1 fill:#2c3e50,stroke:#fff,color:#fff
    style F2 fill:#2c3e50,stroke:#fff,color:#fff
    style F3 fill:#2c3e50,stroke:#fff,color:#fff
    style AI fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style GW fill:#f8f9fa,stroke:#e94560,color:#2c3e50
    style MCP fill:#e94560,stroke:#fff,color:#fff
    style BR fill:#f8f9fa,stroke:#4CAF50,color:#2c3e50
    style SB fill:#f8f9fa,stroke:#4CAF50,color:#2c3e50
    style EX fill:#f8f9fa,stroke:#4CAF50,color:#2c3e50
    style D1 fill:#f8f9fa,stroke:#ff9800,color:#2c3e50
    style R2 fill:#f8f9fa,stroke:#ff9800,color:#2c3e50
    style VZ fill:#f8f9fa,stroke:#ff9800,color:#2c3e50
    style Q fill:#f8f9fa,stroke:#ff9800,color:#2c3e50
    

9. Comparing with other platforms

10. Getting started with Cloudflare Agent Cloud

Here's a simple example of building an AI agent with Project Think:

// wrangler.toml
// name = "my-ai-agent"
// main = "src/index.ts"
// compatibility_date = "2026-04-01"
// [ai]
// binding = "AI"
// [[durable_objects.bindings]]
// name = "AGENT"
// class_name = "ResearchAgent"

import { Think } from "@cloudflare/agents";
import { createWorkersAI } from "@cloudflare/agents/ai";
import { tool } from "ai";
import { z } from "zod";

export class ResearchAgent extends Think<Env> {
  getModel() {
    return createWorkersAI({ binding: this.env.AI })(
      "@cf/google/gemma-4-26b-a4b"
    );
  }

  getTools() {
    return {
      search: tool({
        description: "Search the web for information",
        parameters: z.object({ query: z.string() }),
        execute: async ({ query }) => {
          // Use AI Search or an external API
          const results = await this.env.AI.run(
            "@cf/cloudflare/ai-search",
            { query }
          );
          return results;
        },
      }),
      saveNote: tool({
        description: "Save a note to storage",
        parameters: z.object({
          title: z.string(),
          content: z.string()
        }),
        execute: async ({ title, content }) => {
          // Durable Object SQLite — persists across hibernation
          await this.sql.exec(
            "INSERT INTO notes (title, content, created_at) VALUES (?, ?, ?)",
            title, content, new Date().toISOString()
          );
          return { saved: true };
        },
      }),
    };
  }
}

export default {
  async fetch(request: Request, env: Env): Promise<Response> {
    const id = env.AGENT.idFromName("default");
    const agent = env.AGENT.get(id);
    return agent.fetch(request);
  },
};

Deployment is a single command:

npx wrangler deploy

11. Cloudflare Developer Platform timeline

12. Conclusion

Cloudflare Agent Cloud marks an important shift: from "a place to host websites" to "a place to run AI agents". The combination of V8 isolates (fast, lightweight), Durable Objects (stateful, with hibernation), and a rich storage/AI ecosystem creates a platform few rivals can match on developer experience and operating cost.

The generous free tier — Workers (100K req/day), R2 (10 GB), D1 (5 GB), KV (1 GB), Queues (10K ops/day), Vectorize (5M vectors) — makes it an ideal place to prototype and even run production for small-to-medium AI agent projects at almost zero cost.

References

• Cloudflare Agents Week 2026 — Updates and Announcements
• Project Think: Building the next generation of AI agents on Cloudflare
• Durable Objects in Dynamic Workers: Give each AI-generated app its own database
• Building AI Agents with MCP, Authentication & Authorization, and Durable Objects
• Cloudflare Workers Pricing
• Cloudflare Queues now available on Workers Free plan
• New Workers AI models for text generation and embedding
• Cloudflare Expands its Agent Cloud to Power the Next Generation of Agents