Cloudflare Developer Workshop — Building Durable and Stateful Agents

Building Durable and Stateful Agents

From platform primitives to a deployed AI Agent.

DEVELOPER WORKSHOP

01 / HANDS-ON WORKSHOP

01

Meet the team

DEVELOPER WORKSHOP

MEET THE TEAM

02

Your Cloudflare team.

Anu Razdan

Named Account Executive

anu@cloudflare.com

Terry Mathew

Customer Engineer

tmathew@cloudflare.com

Nikhil Buddhavarapu

CE, Developer & AI

nbuddhavarapu@cloudflare.com

The route / 90 minutes

Understand. Build. Extend.

One application. Small changes. Observable results.

01 Foundations

Workers, bindings, storage, stateful compute.

02 Build the agent

Start with a wired agent. Add state, memory, tools and scheduling.

03 Beyond the lab

Internal app-building with VibeSDK and Void Platform.

15 min primer  ·  55 min build  ·  8 min platform discussion  ·  12 min questions / buffer

DEVELOPER WORKSHOP

THE ROUTE / 90 MINUTES

03

Before you start

Three steps before you build.

Use your favourite editor or coding agent. OpenCode access is linked below.

Sign up for a free Cloudflare account

Local tools

Node.js 24+ and Git. Use the repo’s AGENTS.md with your coding agent.

Accept the lab invite

Create your Cloudflare login first, then accept your workshop lab invite.

DEVELOPER WORKSHOP

BEFORE YOU START

04

dash.cloudflare.com/sign-up

OPENCODE ACCESS

https://opencode.cf.prompt2prod.dev/

Platform primer

A small set of composable primitives.

Compute + storage + AI. Pick the boundary that matches the job.

COMPUTE

STORAGE

AI + CONTROL

Workers

Request handlers + APIs

Durable Objects

Identity + stateful compute

Containers / Sandboxes

Linux execution + isolated code

R2

Object storage + artifacts

D1

Relational application data

KV

Read-heavy configuration

Workers AI

Model inference

AI Gateway

Model traffic + observability

Agents SDK / Workflows

Agent runtime + durable orchestration

DEVELOPER WORKSHOP

PLATFORM PRIMER

05

Workers / the full-stack ecosystem

DEVELOPER WORKSHOP

WORKERS / THE FULL-STACK ECOSYSTEM

06

Workers runs full-stack applications.

Bring your framework. Add server rendering, APIs, assets and platform bindings.

ReactClient UI + Worker API

TanStack StartSSR + server functions

AstroCloudflare adapter

Next.jsvinext / OpenNext paths

vinext

vinextVite-based · beta · check compatibility

SvelteKitadapter-cloudflare

VueVite SPA · Nuxt for SSR

HonoAPIs + middleware

ViteCloudflare Vite plugin

Framework support ≠ identical features. Choose the adapter and test the runtime compatibility.

Bindings / the connective tissue

Everything is a binding.

Your Worker in the center. Capabilities on env. No resource API keys in your code.

DEVELOPER WORKSHOP

BINDINGS / THE CONNECTIVE TISSUE

07

Capabilities on env · configured in Wrangler

● ● ●src/index.ts · all five bindings

export default {
  async fetch(req: Request, env: Env) {
    const flags = await env.CONFIG.get("flags", "json");
    await env.BUCKET.put("hello.txt", "Hello");
    const rows = await env.DB.prepare("SELECT 1").all();
    const ai = await env.AI.run(
      "@cf/zai-org/glm-4.7-flash",
      { messages: [{ role: "user", content: "Hi" }] }
    );
    const room = env.ROOMS.getByName("demo");
    const state = await room.fetch("https://room/state");
    return Response.json({ flags, rows, ai, ok: state.ok });
  }
} satisfies ExportedHandler<Env>;

Wrangler / project configuration

Wrangler: configure, develop, deploy.

Wrangler is Cloudflare’s CLI. wrangler.jsonc defines your Worker, bindings and deployment settings.

DEVELOPER WORKSHOP

WRANGLER / PROJECT CONFIGURATION

08

● ● ●wrangler.jsonc · starter configuration excerpt

{
  "name": "cloudflare-agent-starter",
  "main": "src/server.ts",
  "compatibility_date": "2026-10-19",
  "ai": { "binding": "AI", "remote": true },
  "vars": { "AI_GATEWAY_ID": "<lab-gateway-id>" },
  "durable_objects": { "bindings": [
    { "name": "Assistant", "class_name": "Assistant" }
  ] }
}

A CLI + A CONFIG FILE

Configure

wrangler.jsonc declares the Worker and its bindings.

Develop locally

npm run dev uses Vite + Cloudflare tooling.

Build and deploy

npm run deploy builds and runs Wrangler.

CONFIGURATION   /   Entry point · runtime compatibility · bindings · variables · deployment environments

Use the repo’s complete config, including migrations. Set AI_GATEWAY_ID for your assigned lab account.

Workshop / the starter

DEVELOPER WORKSHOP

WORKSHOP / THE STARTER

09

Start with a working agent.

Four checkpoints. One durable, stateful AI agent.

Streaming chat.
Persistent conversations.
Every model call through AI Gateway.

Agents SDK + Think provide the chat loop, WebSocket sync and message storage. You add the capabilities—not the plumbing.

STARTER REPOSITORY

https://github.com/vnikhilbuddhavarapu/cloudflare-agent-starter

OPENCODE ACCESS

https://opencode.cf.prompt2prod.dev/

ONE INSTANCE NAME → ONE DURABLE OBJECT

Browser

Chat, streaming replies and WebSocket sync.

Assistant · Durable Object

Named compute with its own built-in SQLite.

Agents SDK + Think

Chat loop, transcript storage and tool execution.

AI Gateway → Workers AI

Every model call follows the configured gateway.

Each conversation has its own named agent and SQLite database. Each exercise has a test to turn green.

CP01 / Validated shared settings

DEVELOPER WORKSHOP

CP01 / VALIDATED SHARED SETTINGS

10

One setting. Every tab.

Validated shared settings · add one capability, then prove it.

Add a response style:
concise or detailed.

AGENT STATE   /   setState · validateStateChange · @callable

The browser requests a change. The server validates it, saves it and broadcasts the new state.

● ● ●CHECKPOINT TEST · red → green

npm run test:cp01

CP01 / FOLLOW THE CAPABILITY

Request a style

The browser calls a narrow server method.

Validate server-owned state

Reject direct client writes and invalid values.

Save + broadcast

Every tab on this instance sees the same setting.

PROVE IT

Change style in tab A → tab B updates. Refresh → the setting remains.

CP01 GUIDE

Open checkpoint instructions ↗

CP02 / Bounded memory

DEVELOPER WORKSHOP

CP02 / BOUNDED MEMORY

11

Remember beyond the conversation.

Bounded memory · add one capability, then prove it.

Give the agent memory
that is not chat history.

THINK CONTEXT BLOCK   /   bounded · writable · SQLite-backed

The model writes durable preferences into a context block. A token budget keeps that memory bounded.

● ● ●CHECKPOINT TEST · red → green

npm run test:cp02

CP02 / FOLLOW THE CAPABILITY

Write a preference

The model uses set_context to update memory.

Enforce the budget

Writes beyond the 1,000-token limit are rejected.

Recall on a later turn

SQLite memory survives clearing chat history.

PROVE IT

Save a preference → clear the chat → ask again. The preference survives.

CP02 GUIDE

Open checkpoint instructions ↗

CP03 / Typed note tools

DEVELOPER WORKSHOP

CP03 / TYPED NOTE TOOLS

12

Useful tools. Checked inputs.

Typed note tools · add one capability, then prove it.

Let the model save
and list durable notes.

AI SDK tool()   /   Zod schemas · SQLite · toolCallId

Validate the inputs, store the record and return structured results. A retry with the same call ID returns the original note.

● ● ●CHECKPOINT TEST · red → green

npm run test:cp03

CP03 / FOLLOW THE CAPABILITY

Validate with Zod

Trim and bound the note title and body.

Save once in SQLite

Use toolCallId as the operation key.

Return a tool result

The chat renders the input and structured result.

PROVE IT

Run the direct tool test. Then ask the model: a tool card and one saved note appear.

CP03 GUIDE

Open checkpoint instructions ↗

CP04 / Durable reminders

DEVELOPER WORKSHOP

CP04 / DURABLE REMINDERS

13

Close the tab. Keep the promise.

Durable reminders · add one capability, then prove it.

Schedule reminders
that outlive the browser.

this.schedule()   /   durable alarm · idempotent completion

A persisted schedule wakes the Durable Object when due. Completion is stored before clients are notified—and is safe to repeat.

● ● ●CHECKPOINT TEST · red → green

npm run test:cp04

CP04 / FOLLOW THE CAPABILITY

Register durable work

Validate the delay; this.schedule() takes seconds.

Wake the named agent

The alarm fires even with no connected tabs.

Complete idempotently

Persist pending → completed once; ignore repeats.

PROVE IT

Ask for 20 seconds → close every tab → reopen after it’s due. Status: completed.

CP04 GUIDE

Open checkpoint instructions ↗

Beyond the workshop / agent capabilities

DEVELOPER WORKSHOP

BEYOND THE WORKSHOP / AGENT CAPABILITIES

14

Your agent can do more.

New channels. Longer-running work. Specialist agents.

CHANNELS + HUMAN OVERSIGHT

MORE WAYS TO GET WORK DONE

Email

Inbound mail + replies

Voice

Streaming speech · beta

Video

RealtimeKit / WebRTC

HITL

Human-in-the-loop approvals

Fibers

Checkpoint + recover work

Sub-agents

Delegate with isolated state

MCP tools

Connect external capabilities

Workflows

Durable steps + retries

Explore the Agents SDK primitives ↗

CP05 / Deploy and submit

DEVELOPER WORKSHOP

CP05 / DEPLOY AND SUBMIT

15

Ship it. Share what you built.

Deploy your agent, then submit it with a short description of what you built.

Deploy

Set AI_GATEWAY_ID in wrangler.jsonc
Run npm run deploy to get your workers.dev URL
Open it, chat, and check your checkpoints work live

Submit

Your deployed URL
GitHub repo or code link, if shareable
A 2–3 sentence description of your agent
Checkpoints or stretch goals completed

CP05 GUIDE

Open deploy and submit instructions ↗

PROJECT SUBMISSION FORMhttps://forms.gle/B9PYQHPZwYJAsxFq9

No login by default: protect longer-lived deployments with Cloudflare Access or AI Gateway rate limits.

DEVELOPER WORKSHOP

THANK YOU

16

Thank you!