Changelog

Agents connected to Model Context Protocol (MCP) servers with addMcpServer can now handle elicitation ↗ requests.

Elicitation lets an MCP server request user input while it handles a tool call. Form mode collects structured, non-sensitive data. URL mode asks for consent before opening an out-of-band flow, such as third-party authorization or payment.

sequenceDiagram
    participant User
    participant Agent as Agent (MCP client)
    participant Server as MCP server
    participant Browser

    Server->>Agent: elicitation/create
    Agent->>User: Show server, reason, and input or URL
    User->>Agent: Submit, open, decline, or cancel
    Agent->>Browser: Open URL after consent (URL mode)
    Agent->>Server: accept, decline, or cancel
    Server-->>Agent: Optional URL completion notification

Register a handler for each mode your Agent supports in onStart():

import { Agent } from "agents";

export class MyAgent extends Agent {
	onStart() {
		this.mcp.configureElicitationHandlers({
			form: (request, serverId) => this.forwardToUser(request, serverId),
			url: (request, serverId) => this.forwardToUser(request, serverId),
		});
	}

	forwardToUser(request, serverId) {
		// Show the request in your UI and resolve after the user responds.
		throw new Error(
			`Implement elicitation for ${serverId}: ${request.params.message}`,
		);
	}
}

import { Agent } from "agents";
import type { ElicitRequest, ElicitResult } from "agents/mcp";

export class MyAgent extends Agent<Env> {
	onStart() {
		this.mcp.configureElicitationHandlers({
			form: (request, serverId) => this.forwardToUser(request, serverId),
			url: (request, serverId) => this.forwardToUser(request, serverId),
		});
	}

	private forwardToUser(
		request: ElicitRequest,
		serverId: string,
	): Promise<ElicitResult> {
		// Show the request in your UI and resolve after the user responds.
		throw new Error(
			`Implement elicitation for ${serverId}: ${request.params.message}`,
		);
	}
}

Connections advertise only the modes with configured handlers. An Agent without handlers advertises no elicitation capability, which lets the server use its fallback. The SDK stores the advertised modes with each MCP server registration so they survive Durable Object hibernation. Callback functions remain in memory and reattach when onStart() runs.

For implementation details and a browser forwarding pattern, refer to MCP client elicitation. The mcp-client ↗ and mcp-elicitation ↗ examples implement both sides.

To update to this release:

								
									
									npm
									 i agents@latest
								
							

								
									
									yarn
									 add agents@latest
								
							

								
									
									pnpm
									 add agents@latest
								
							

								
									
									bun
									 add agents@latest
								
							

The latest release of the Agents SDK ↗ makes it easier to run long work in the background, drive turns through one entry point, and keep chat agents working through deploys, evictions, and reconnects.

This release adds first-class detached (background) sub-agent runs with live progress and durable milestones, a single runTurn turn-admission entry point, and a large round of recovery and reliability fixes that continue converging @cloudflare/think and @cloudflare/ai-chat onto one model.

runAgentTool can now dispatch a sub-agent without blocking the calling turn. A detached run returns a handle immediately and is owned by a durable, eviction-surviving backbone instead of being abandoned when the dispatching turn ends.

class OrdersAgent extends Think {
	async startImport(input) {
		// Fire-and-forget, or wire a durable completion callback
		// (by method name, like schedule()):
		await this.runAgentTool(ImportAgent, {
			input,
			detached: { onFinish: "onImportDone", maxBudgetMs: 60 * 60 * 1000 },
		});
	}

	// result.status: "completed" | "error" | "aborted" | "interrupted"
	async onImportDone(run, result) {}
}

class OrdersAgent extends Think {
	async startImport(input) {
		// Fire-and-forget, or wire a durable completion callback
		// (by method name, like schedule()):
		await this.runAgentTool(ImportAgent, {
			input,
			detached: { onFinish: "onImportDone", maxBudgetMs: 60 * 60 * 1000 },
		});
	}

	// result.status: "completed" | "error" | "aborted" | "interrupted"
	async onImportDone(run, result) {}
}

Highlights:

• Durable, exactly-once-on-the-happy-path completion via a warm fast path plus a self-scheduling reconcile backbone that survives eviction and deploys.
• Bounded. An absolute maxBudgetMs ceiling (default 24h) and cancelAgentTool(runId) keep abandoned runs from holding a concurrency slot forever.
• detached: { notify: true } lets a finished background run inject a message back into the chat so the model reacts to the result — no hand-wired onFinish needed.

Sub-agents can also report mid-run progress that rides their own turn stream back to the parent's connected clients:

// Inside the child sub-agent:
await this.reportProgress({
	fraction: 0.6,
	phase: "deploying",
	message: "Generating menu page…",
});

// Inside the child sub-agent:
await this.reportProgress({
	fraction: 0.6,
	phase: "deploying",
	message: "Generating menu page…",
});

Progress surfaces on AgentToolRunState.progress via useAgentToolEvents, so a background-runs tray can render a live bar without drilling in, and the latest snapshot is persisted for inspection after eviction. Naming a milestone promotes a signal to a durable, replayable row, and detached: { onMilestones } can surface a milestone as a synthetic chat message ("narrate" for a cheap status line, or "react" to drive a model turn).

@cloudflare/think adds a public runTurn(options) facade that unifies turn admission behind a single mode:

await this.runTurn({ mode: "wait", messages }); // saveMessages / continueLastTurn
await this.runTurn({ mode: "submit", messages }); // durable submitMessages
await this.runTurn({ mode: "stream", messages }); // chat()

await this.runTurn({ mode: "wait", messages }); // saveMessages / continueLastTurn
await this.runTurn({ mode: "submit", messages }); // durable submitMessages
await this.runTurn({ mode: "stream", messages }); // chat()

stream mode accepts array and function inputs to match wait mode, and all entry points now route through a shared internal admission path that throws a clear error on nested blocking admissions that previously could deadlock.

A large part of this release continues hardening recovery and converging @cloudflare/think and @cloudflare/ai-chat onto one model:

• Stream stall watchdog. AIChatAgent can detect and recover from a hung model/transport stream via the opt-in chatStreamStallTimeoutMs watchdog. With chatRecovery enabled the stall routes into the same bounded-recovery machinery a deploy or eviction uses; otherwise it surfaces as a terminal stream error so the spinner clears.
• Interrupted tool-call repair. AIChatAgent now repairs a transcript with a dead server-tool call before re-entering inference (parity with @cloudflare/think), so a recovered turn no longer fails with AI_MissingToolResultsError. An overridable repairInterruptedToolPart(part) hook lets apps customize the repaired shape.
• Stuck status after reconnect. Fixed AI SDK status getting stuck when a reconnect races a turn that has been accepted but has not started streaming yet, so the UI now renders the in-flight turn instead of settling on ready.
• Live "recovering…" on connect. AIChatAgent now replays the recovering status to a client that connects mid-recovery, so useAgentChat's isRecovering reflects in-progress recovery immediately instead of appearing frozen.
• Terminal connection failures. The client stops reconnecting on terminal WebSocket close events and exposes them via connectionError / onConnectionError on AgentClient, useAgent, and useAgentChat.
• Agent-tool child recovery. A healthy long-running sub-agent run is no longer abandoned as interrupted after a deploy (both @cloudflare/think and AIChatAgent).
• Workflows from sub-agent facets. Agent Workflows can now start from sub-agent facets, with callbacks and Workflow RPC routed back to the originating facet.
• Plus forward-progress crediting convergence, broadcast-first give-up ordering, an event-driven auto-continuation barrier, and structured row-size compaction in AIChatAgent.

• Shared chat React core. A new agents/chat/react entry exposes useAgentChat, transport helpers, and shared wire types, with syncMessagesToServer for server-authoritative transcript storage. @cloudflare/think/react and @cloudflare/ai-chat/react are now thin wrappers over it.
• Optional ai peer. The root agents and @cloudflare/codemode runtimes no longer reference AI SDK types, so they bundle without ai / zod installed; AI-specific entry points still require the peer when imported. just-bash likewise moves to an optional peer used only by the skills bash runner.
• Code Mode. The default DynamicWorkerExecutor timeout increases from 30s to 60s, executions now dispose the dynamically-loaded Worker and its RPC stub after each run (fixing a flaky isolate-shutdown assertion), connector imports are cleaned up, and the outer MCP tool-call context is passed to openApiMcpServer request callbacks.
• Voice. Voice turns now support AI SDK fullStream responses (and warn when textStream is used).
• MCP. McpAgent server-to-client requests can now be sent from callbacks that do not inherit the agent's async context, including callbacks reached through Worker Loader RPC.
• Experimental: server actions and channels. This release lays groundwork for guarded server actions (action() / getActions() with a durable replay ledger and approvals) and a unified channels surface (configureChannels(), deliverNotice()). Both are experimental and their APIs may change, so we don't recommend depending on them yet.

To update to the latest version:

								
									
									npm
									 i agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest @cloudflare/codemode@latest @cloudflare/voice@latest
								
							

								
									
									yarn
									 add agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest @cloudflare/codemode@latest @cloudflare/voice@latest
								
							

								
									
									pnpm
									 add agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest @cloudflare/codemode@latest @cloudflare/voice@latest
								
							

								
									
									bun
									 add agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest @cloudflare/codemode@latest @cloudflare/voice@latest
								
							

Refer to the Think documentation, Code Mode documentation, and Agents documentation for more information.

The latest release of the Agents SDK ↗ makes it easier to build agents that can safely interact with real systems and keep working through interruptions.

Agents can now browse websites through Browser Run, write code against external tools through Code Mode, use client-provided tools when delegating to Think sub-agents, and recover more reliably from deploys, Durable Object evictions, and connection churn.

Agents can now use Browser Run through a single durable browser_execute tool. Instead of choosing from a fixed list of actions, the model writes code against the Chrome DevTools Protocol (CDP) and can inspect pages, capture screenshots, read rendered content, debug frontend behavior, and interact with live browser sessions.

const browserTools = createBrowserTools({
	ctx: this.ctx,
	browser: this.env.BROWSER,
	loader: this.env.LOADER,
	session: { mode: "dynamic" },
});

const browserTools = createBrowserTools({
	ctx: this.ctx,
	browser: this.env.BROWSER,
	loader: this.env.LOADER,
	session: { mode: "dynamic" },
});

Browser sessions can be one-time, reused, or promoted from one-time to persistent during a run. This is useful when an agent needs a human to log in, complete MFA, or approve a sensitive action. The run can pause, keep the same tabs and cookies, and resume after approval.

The browser tools also add Live View URLs, optional session recording, and quick actions such as browser_markdown, browser_extract, browser_links, and browser_scrape for one-shot browsing tasks.

Code Mode now uses createCodemodeRuntime, connectors, and a durable execution log. This lets you give a model one codemode tool instead of a large prompt full of tool definitions. The model can discover the capabilities it needs, write code against typed globals, and reuse saved snippets.

const runtime = createCodemodeRuntime({
	ctx: this.ctx,
	executor: new DynamicWorkerExecutor({ loader: this.env.LOADER }),
	connectors: [new GithubConnector(this.ctx, this.env, connection)],
});

const result = streamText({
	model,
	messages,
	tools: { codemode: runtime.tool() },
});

const runtime = createCodemodeRuntime({
	ctx: this.ctx,
	executor: new DynamicWorkerExecutor({ loader: this.env.LOADER }),
	connectors: [new GithubConnector(this.ctx, this.env, connection)],
});

const result = streamText({
	model,
	messages,
	tools: { codemode: runtime.tool() },
});

When the code reaches an approval-gated action, the runtime pauses execution and returns a pending approval. After approval, completed calls replay from the durable log, the approved action runs, and the same code continues. This makes it practical to build agents that create issues, update external systems, or perform other side effects without custom pause-and-resume logic for every tool.

Think sub-agents can now use client-defined tools over the RPC chat() path. A parent agent can pass tool schemas with clientTools and resolve tool calls through onClientToolCall. This lets delegated agents use caller-provided capabilities without requiring a browser WebSocket.

await child.chat(message, callback, {
	signal,
	clientTools: [
		{
			name: "get_user_timezone",
			description: "Get the caller's timezone",
			parameters: { type: "object" },
		},
	],
	onClientToolCall: async ({ toolName, input }) => {
		return runClientTool(toolName, input);
	},
});

await child.chat(message, callback, {
	signal,
	clientTools: [
		{
			name: "get_user_timezone",
			description: "Get the caller's timezone",
			parameters: { type: "object" },
		},
	],
	onClientToolCall: async ({ toolName, input }) => {
		return runClientTool(toolName, input);
	},
});

Think Workflows also improve step.prompt(). A prompt step now runs a full agentic turn before returning structured output, so the agent can call tools before producing the typed result. This makes Workflow steps more useful for durable triage, research, and approval flows.

The unified Think execute tool can also include cdp.* browser capabilities alongside state.* and tools.* when Browser Run is bound.

Voice clients can route assistant audio to a specific output device. Use outputDeviceId with useVoiceAgent, or call client.setOutputDevice() from the framework-agnostic client.

const voice = useVoiceAgent({
	agent: "MyVoiceAgent",
	outputDeviceId: selectedSpeakerId,
});

const voice = useVoiceAgent({
	agent: "MyVoiceAgent",
	outputDeviceId: selectedSpeakerId,
});

Browsers without speaker-selection support continue playing through the default output device and report a non-fatal outputDeviceError.

This release includes several fixes for production agents:

• useAgent and AgentClient handle WebSocket replacement more reliably during reconnects and configuration changes.
• Chat stream replay is more reliable after reconnects, deploys, and provider errors.
• Fiber recovery continues across multi-pass scans and backs off when recovery hooks keep failing.
• Agent teardown continues even when the request that started teardown is canceled.
• Large session histories use byte-budgeted reads to reduce memory pressure during startup.

To update to the latest version:

								
									
									npm
									 i agents@latest @cloudflare/think@latest @cloudflare/codemode@latest @cloudflare/ai-chat@latest @cloudflare/voice@latest
								
							

								
									
									yarn
									 add agents@latest @cloudflare/think@latest @cloudflare/codemode@latest @cloudflare/ai-chat@latest @cloudflare/voice@latest
								
							

								
									
									pnpm
									 add agents@latest @cloudflare/think@latest @cloudflare/codemode@latest @cloudflare/ai-chat@latest @cloudflare/voice@latest
								
							

								
									
									bun
									 add agents@latest @cloudflare/think@latest @cloudflare/codemode@latest @cloudflare/ai-chat@latest @cloudflare/voice@latest
								
							

Refer to the Code Mode documentation, Browser tools documentation, Think tools documentation, and Voice documentation for more information.

We are excited to announce GLM-5.2 on Workers AI, Z.ai's flagship agentic coding model.

@cf/zai-org/glm-5.2 is a text generation model built for agentic coding workflows. With function calling and reasoning support, it can handle long codebases, multi-step planning, and tool-augmented agents.

Key features and use cases:

• Agentic coding: Designed for autonomous coding tasks, long-horizon planning, and complex software engineering workflows
• Large context window: GLM-5.2 supports up to a 1,048,576 token context window. Workers AI is launching the model with a 262,144 token context window and plans to increase this in the future
• Function calling: Build agents that invoke tools and APIs across multiple conversation turns
• Reasoning: Tackles complex problem-solving and step-by-step reasoning tasks

Use GLM-5.2 through the Workers AI binding (env.AI.run()), the REST API at /run or /v1/chat/completions, or AI Gateway.

Pricing is available on the model page or pricing page.

The latest release of the Agents SDK ↗ adds four new ways to build with @cloudflare/think: on-demand Agent Skills, chat messengers (starting with Telegram), declarative scheduled tasks, and durable reasoning steps inside Workflows. This release also significantly hardens durable chat recovery, so turns reliably ride through deploys, evictions, and stalled model streams in production.

Give an agent a catalog of on-demand instructions, resources, and scripts. A skill source adds a catalog to the system prompt, and the model activates a skill only when a task matches — so a large library of capabilities does not bloat every prompt.

import { Think, skills } from "@cloudflare/think";
import bundledSkills from "agents:skills";

export class SkillsAgent extends Think {
	getSkills() {
		return [
			bundledSkills,
			skills.r2(this.env.SKILLS_BUCKET, { prefix: "skills/" }),
		];
	}
}

import { Think, skills } from "@cloudflare/think";
import bundledSkills from "agents:skills";

export class SkillsAgent extends Think<Env> {
	getSkills() {
		return [
			bundledSkills,
			skills.r2(this.env.SKILLS_BUCKET, { prefix: "skills/" }),
		];
	}
}

The agents:skills import bundles a local ./skills directory through the Agents Vite plugin (one directory per skill, each with a SKILL.md). Skills can also load from R2 or a manifest. When skills are available, Think exposes activate_skill, read_skill_resource, and an optional run_skill_script tool. Skill loading is resilient: a duplicate or failing source is skipped with a warning instead of breaking the agent.

Agent Skills are experimental, and script execution in particular is early. The API may change in a future release. We would love your feedback — tell us what you are building and what is missing in the Agents repository ↗.

Connect a Think agent directly to a chat platform. Think owns the webhook route, conversation routing, durable reply fiber, and streamed delivery back to the provider. Telegram ships as the first provider.

import { Think } from "@cloudflare/think";
import {
	defineMessengers,
	ThinkMessengerStateAgent,
} from "@cloudflare/think/messengers";
import telegramMessenger from "@cloudflare/think/messengers/telegram";

export { ThinkMessengerStateAgent };

export class SupportAgent extends Think {
	getMessengers() {
		return defineMessengers({
			telegram: telegramMessenger({
				token: this.env.TELEGRAM_BOT_TOKEN,
				userName: "support_bot",
				secretToken: this.env.TELEGRAM_WEBHOOK_SECRET_TOKEN,
			}),
		});
	}
}

import { Think } from "@cloudflare/think";
import {
	defineMessengers,
	ThinkMessengerStateAgent,
} from "@cloudflare/think/messengers";
import telegramMessenger from "@cloudflare/think/messengers/telegram";

export { ThinkMessengerStateAgent };

export class SupportAgent extends Think<Env> {
	getMessengers() {
		return defineMessengers({
			telegram: telegramMessenger({
				token: this.env.TELEGRAM_BOT_TOKEN,
				userName: "support_bot",
				secretToken: this.env.TELEGRAM_WEBHOOK_SECRET_TOKEN,
			}),
		});
	}
}

Each Chat SDK thread maps to its own Think sub-agent by default, so group chats and direct messages do not share memory. Multiple bots, custom conversation routing, and custom providers are all supported.

Declare recurring, timezone-aware prompts and handlers with a typed domain-specific language (DSL). Think reconciles the declarations on startup and re-arms the next occurrence after each run, backed by durable idempotent submissions.

import { Think, defineScheduledTasks } from "@cloudflare/think";

export class DigestAgent extends Think {
	getScheduledTasks() {
		return defineScheduledTasks({
			weeklyCommitReport: {
				schedule: "every week on monday at 09:00",
				prompt:
					"Compile my GitHub commits for the last week and summarize them.",
			},
			workout: {
				schedule: "every day at 08:00 in Europe/London",
				prompt: "Start my workout.",
			},
		});
	}
}

import { Think, defineScheduledTasks } from "@cloudflare/think";

export class DigestAgent extends Think<Env> {
	getScheduledTasks() {
		return defineScheduledTasks({
			weeklyCommitReport: {
				schedule: "every week on monday at 09:00",
				prompt:
					"Compile my GitHub commits for the last week and summarize them.",
			},
			workout: {
				schedule: "every day at 08:00 in Europe/London",
				prompt: "Start my workout.",
			},
		});
	}
}

Run a model-driven reasoning step inside a Cloudflare Workflow with ThinkWorkflow and step.prompt(), with durable typed structured output, long waits, and approval gates.

import { z } from "zod";
import { ThinkWorkflow } from "@cloudflare/think/workflows";

const draftSchema = z.object({
	title: z.string(),
	summary: z.string(),
	labels: z.array(z.string()),
});

export class TriageWorkflow extends ThinkWorkflow {
	async run(event, step) {
		const draft = await step.prompt("triage-issue", {
			prompt: `Triage issue #${event.payload.issueNumber}`,
			output: draftSchema,
			timeout: "3 days",
		});

		await step.do("apply-labels", async () => {
			await this.agent.applyLabels(draft.labels);
		});
	}
}

import { z } from "zod";
import { ThinkWorkflow } from "@cloudflare/think/workflows";
import type { ThinkWorkflowStep } from "@cloudflare/think/workflows";
import type { AgentWorkflowEvent } from "agents/workflows";

const draftSchema = z.object({
	title: z.string(),
	summary: z.string(),
	labels: z.array(z.string()),
});

export class TriageWorkflow extends ThinkWorkflow<TriageAgent, Params> {
	async run(event: AgentWorkflowEvent<Params>, step: ThinkWorkflowStep) {
		const draft = await step.prompt("triage-issue", {
			prompt: `Triage issue #${event.payload.issueNumber}`,
			output: draftSchema,
			timeout: "3 days",
		});

		await step.do("apply-labels", async () => {
			await this.agent.applyLabels(draft.labels);
		});
	}
}

Durable chat turns have always been designed to survive a mid-turn deploy or Durable Object eviction. This release is a major hardening pass on that machinery for production.

• Better recovery during deploys. Turns now ride through continuous deploys and evictions without losing completed work or re-running tools that already ran.
• A live "recovering…" signal. useAgentChat exposes a new isRecovering flag, so a recovering turn shows progress instead of looking frozen. Most UIs render isStreaming || isRecovering as "busy".
• Stalled streams recover. Set chatStreamStallTimeoutMs to route a hung provider stream into the same recovery path instead of leaving an infinite spinner.
• Sub-agents re-attach. On parent recovery, an in-flight agentTool() child is re-attached to its result rather than abandoned and re-run, so long-running children no longer lose work under deploys.

• Resumable streams — In-flight tool calls over Server-Sent Events (SSE) survive a dropped connection. Clients reconnect with Last-Event-ID and replay anything they missed.
• Readable server IDs — addMcpServer accepts an optional id, so tools surface as readable keys (for example tool_github_create_pull_request) instead of opaque connection IDs.
• Better handling of concurrent requests — Overlapping JSON-RPC requests are now correctly correlated to their responses across the HTTP and RPC transports.

• Compaction — A Session's tokenCounter now also drives the compaction boundary decision ("what to compress"), not just the fire/no-fire trigger.
• @cloudflare/worker-bundler — Adds a virtualModules option to createWorker to provide in-memory module source during bundling.
• Client-tool continuations — Parallel tool results now coalesce into a single continuation, immediate resume requests attach to the pending continuation, and server-side needsApproval continuations resume reliably after approval.

To update to the latest version:

								
									
									npm
									 i agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest
								
							

								
									
									yarn
									 add agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest
								
							

								
									
									pnpm
									 add agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest
								
							

								
									
									bun
									 add agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest
								
							

Refer to the Agents API reference and Chat agents documentation for more information.

Sandboxes can expose a service running inside the container on a public preview URL through the sandbox.tunnels namespace. The SDK uses cloudflared inside the sandbox so you can share a running service without configuring exposePort() or a custom domain.

By default, sandbox.tunnels.get(port) creates a quick tunnel ↗ on a zero-config *.trycloudflare.com URL — no Cloudflare account, DNS record, or custom domain required. This is perfect for quick development and for .workers.dev deployments.

import { getSandbox } from "@cloudflare/sandbox";

const sandbox = getSandbox(env.Sandbox, "my-sandbox");
await sandbox.startProcess("python -m http.server 8080");

const tunnel = await sandbox.tunnels.get(8080);
console.log(tunnel.url); // → https://random-words-here.trycloudflare.com

import { getSandbox } from "@cloudflare/sandbox";

const sandbox = getSandbox(env.Sandbox, "my-sandbox");
await sandbox.startProcess("python -m http.server 8080");

const tunnel = await sandbox.tunnels.get(8080);
console.log(tunnel.url); // → https://random-words-here.trycloudflare.com

For more control you can create a named tunnel through sandbox.tunnels.get(port, { name }). A named tunnel binds a hostname (<name>.<your-zone>) backed by a Cloudflare Tunnel and a CNAME record on your zone resulting in something like https://my-app-preview.example.com ↗.

Unlike quick tunnels, which generate a new random URL each time, a named tunnel produces a persistent URL that survives container restarts. This makes named tunnels suitable for production use cases where you want control over the tunnel and it's origin.

const tunnel = await sandbox.tunnels.get(8080, { name: "my-app-preview" });
console.log(tunnel.url); // → https://my-app-preview.example.com

const tunnel = await sandbox.tunnels.get(8080, { name: "my-app-preview" });
console.log(tunnel.url); // → https://my-app-preview.example.com

Calling sandbox.destroy() tears down the Cloudflare Tunnel and the associated DNS record alongside the container, so you do not leave dangling tunnels or records behind.

To update to the latest version:

								
									
									npm
									 i @cloudflare/sandbox@latest
								
							

								
									
									yarn
									 add @cloudflare/sandbox@latest
								
							

								
									
									pnpm
									 add @cloudflare/sandbox@latest
								
							

								
									
									bun
									 add @cloudflare/sandbox@latest
								
							

For full API details, refer to the Sandbox tunnels reference.

The latest release of the Agents SDK ↗ brings more reliable chat recovery, fixes Agent state synchronization during reconnects, adds durable submissions for Think, exposes routing retry configuration, and adds connection control for Voice agents.

@cloudflare/ai-chat now keeps server turns running when a browser or client stream is interrupted. This is useful for long-running AI responses where users refresh the page, close a tab, or temporarily lose connection. Calling stop() still cancels the server turn.

Set cancelOnClientAbort: true if browser or client aborts should also cancel the server turn:

const chat = useAgentChat({
	agent: "assistant",
	name: "user-123",
	cancelOnClientAbort: true,
});

const chat = useAgentChat({
	agent: "assistant",
	name: "user-123",
	cancelOnClientAbort: true,
});

Notable bug fixes:

• Chat stream resume negotiation no longer throws when replay races with a closed WebSocket connection.
• Recovered chat continuations no longer leave useAgentChat stuck in a streaming state when the original socket disconnects before a terminal response.
• Approval auto-continuation preserves reasoning parts and persists continuation reasoning in the final message.
• isServerStreaming now resets correctly when a resumed stream moves from the fallback observer path to a transport-owned stream.

agents@0.12.4 prevents duplicate initial state frames during WebSocket connection setup. This avoids stale initial state messages overwriting state updates already sent by the client.

Agent recovery is also more reliable when tool calls span a Durable Object restart. Recovery now defers user finish hooks until after agent startup and isolates hook failures, so one failed hook does not block other recovered runs from finalizing.

getAgentByName() now supports routingRetry for transient Durable Object routing failures:

import { getAgentByName } from "agents";

const agent = await getAgentByName(env.AssistantAgent, "user-123", {
	routingRetry: {
		maxAttempts: 3,
	},
});

import { getAgentByName } from "agents";

const agent = await getAgentByName(env.AssistantAgent, "user-123", {
	routingRetry: {
		maxAttempts: 3,
	},
});

@cloudflare/think now supports durable programmatic submissions. submitMessages() provides durable acceptance, idempotent retries, status inspection, cancellation, and cleanup for server-driven turns that should continue after the caller returns.

Think.chat() RPC turns now run inside chat recovery fibers and persist their stream chunks. Interrupted sub-agent turns can recover partial output instead of starting over.

ChatOptions.tools has been removed from the TypeScript API. Define durable tools on the child agent or use agent tools for orchestration. Runtime options.tools values passed by legacy callers are ignored with a warning.

@cloudflare/think no longer applies pruneMessages({ toolCalls: "before-last-2-messages" }) to model context by default. The previous default could strip client-side tool results from longer multi-turn flows.

truncateOlderMessages still runs as before, so context cost remains bounded. Subclasses that relied on the old aggressive pruning can opt back in from beforeTurn:

import { Think } from "@cloudflare/think";
import { pruneMessages } from "ai";

export class MyAgent extends Think {
	beforeTurn(ctx) {
		return {
			messages: pruneMessages({
				messages: ctx.messages,
				toolCalls: "before-last-2-messages",
			}),
		};
	}
}

import { Think } from "@cloudflare/think";
import { pruneMessages } from "ai";

export class MyAgent extends Think<Env> {
	beforeTurn(ctx) {
		return {
			messages: pruneMessages({
				messages: ctx.messages,
				toolCalls: "before-last-2-messages",
			}),
		};
	}
}

@cloudflare/voice adds an enabled option to useVoiceAgent. React apps can now delay creating and connecting a VoiceClient until prerequisites such as capability tokens are ready.

const voice = useVoiceAgent({
	agent: "MyVoiceAgent",
	enabled: Boolean(token),
});

const voice = useVoiceAgent({
	agent: "MyVoiceAgent",
	enabled: Boolean(token),
});

This release also fixes Workers AI speech-to-text session edge cases and withVoice text streaming from AI SDK textStream responses.

• Streamable HTTP routing — Server-to-client requests now route through the originating POST stream when no standalone SSE stream is available.
• Structured tool output — Tool output shapes are preserved when truncating older messages or oversized persisted rows.
• Non-chat Think tool steps — Think agent-tool children can complete without emitting assistant text and can return structured output through getAgentToolOutput.
• Sub-agent schedules — Stale sub-agent schedule rows are pruned when their owning facet registry entry no longer exists.
• @cloudflare/codemode — Adds a browser-safe export with an iframe sandbox executor and resolves OpenAPI specs inside the sandbox to avoid Worker Loader RPC size limits.

To update to the latest version:

npm i agents@latest @cloudflare/ai-chat@latest @cloudflare/think@latest @cloudflare/voice@latest

Refer to the Agents API reference and Chat agents documentation for more information.

We are excited to announce two major capability upgrades for Agent Lee, the AI co-pilot built directly into the Cloudflare dashboard. Agent Lee is designed to understand your specific account configuration, and with this release, it moves from a passive advisor to an active assistant that can help you manage your infrastructure and visualize your data through natural language.

Agent Lee can now perform changes on your behalf across your Cloudflare account. Whether you need to update DNS records, modify SSL/TLS settings, or configure Workers routes, you can simply ask.

To ensure security and accuracy, every write operation requires explicit user approval. Before any change is committed, Agent Lee will present a summary of the proposed action in plain language. No action is taken until you select Confirm, and this approval requirement is enforced at the infrastructure level to prevent unauthorized changes.

Example requests:

• "Add an A record for blog.example.com pointing to 192.0.2.10."
• "Enable Always Use HTTPS on my zone."
• "Set the SSL mode for example.com to Full (strict)."

Understanding your traffic and security trends is now as easy as asking a question. Agent Lee now features Generative UI, allowing it to render inline charts and structured data visualizations directly within the chat interface using your actual account telemetry.

Example requests:

• "Show me a chart of my traffic over the last 7 days."
• "What does my error rate look like for the past 24 hours?"
• "Graph my cache hit rate for example.com this week."

These features are currently available in Beta for all users on the Free plan. To get started, log in to the Cloudflare dashboard ↗ and select Ask AI in the upper right corner.

To learn more about how to interact with your account using AI, refer to the Agent Lee documentation.

Outbound Workers for Sandboxes and Containers now support zero-trust credential injection, TLS interception, allow/deny lists, and dynamic per-instance egress policies. These features give platforms running agentic workloads full control over what leaves the sandbox, without exposing secrets to untrusted workloads, like user-generated code or coding agents.

Because outbound handlers run in the Workers runtime, outside the sandbox, they can hold secrets the sandbox never sees. A sandboxed workload can make a plain request, and credentials are transparently attached before a request is forwarded upstream.

For instance, you could run an agent in a sandbox and ensure that any requests it makes to Github are authenticated. But it will never be able to access the credentials:

export class MySandbox extends Sandbox {}

MySandbox.outboundByHost = {
	"github.com": (request: Request, env: Env, ctx: OutboundHandlerContext) => {
		const requestWithAuth = new Request(request);
		requestWithAuth.headers.set("x-auth-token", env.SECRET);
		return fetch(requestWithAuth);
	},
};

You can easily inject unique credentials for different instances by using ctx.containerId:

MySandbox.outboundByHost = {
	"my-internal-vcs.dev": async (
		request: Request,
		env: Env,
		ctx: OutboundHandlerContext,
	) => {
		const authKey = await env.KEYS.get(ctx.containerId);

		const requestWithAuth = new Request(request);
		requestWithAuth.headers.set("x-auth-token", authKey);
		return fetch(requestWithAuth);
	},
};

No token is ever passed into the sandbox. You can rotate secrets in the Worker environment and every request will pick them up immediately.

Outbound Workers now intercept HTTPS traffic. A unique ephemeral certificate authority (CA) and private key are created for each sandbox instance. The CA is placed into the sandbox and trusted by default. The ephemeral private key never leaves the container runtime sidecar process and is never shared across instances.

With TLS interception active, outbound Workers can act as a transparent proxy for both HTTP and HTTPS traffic.

Easily filter outbound traffic with allowedHosts and deniedHosts. When allowedHosts is set, it becomes a deny-by-default allowlist. Both properties support glob patterns.

export class MySandbox extends Sandbox {
	allowedHosts = ["github.com", "npmjs.org"];
}

Define named outbound handlers then apply or remove them at runtime using setOutboundHandler() or setOutboundByHost(). This lets you change egress policy for a running sandbox without restarting it.

export class MySandbox extends Sandbox {}

MySandbox.outboundHandlers = {
	allowHosts: async (req: Request, env: Env, ctx: OutboundHandlerContext ) => {
		const url = new URL(req.url);
		if (ctx.params.allowedHostnames.includes(url.hostname)) {
			return fetch(req);
		}
		return new Response(null, { status: 403 });
	},

	noHttp: async () => {
		return new Response(null, { status: 403 });
	},
};

Apply handlers programmatically from your Worker:

const sandbox = getSandbox(env.Sandbox, userId);

// Open network for setup
await sandbox.setOutboundHandler("allowHosts", {
	allowedHostnames: ["github.com", "npmjs.org"],
});
await sandbox.exec("npm install");

// Lock down after setup
await sandbox.setOutboundHandler("noHttp");

Handlers accept params, so you can customize behavior per instance without defining separate handler functions.

Upgrade to @cloudflare/containers@0.3.0 or @cloudflare/sandbox@0.8.9 to use these features.

For more details, refer to Sandbox outbound traffic and Container outbound traffic.

The latest release of the Agents SDK ↗ exposes agent state as a readable property, prevents duplicate schedule rows across Durable Object restarts, brings full TypeScript inference to AgentClient, and migrates to Zod 4.

Both useAgent (React) and AgentClient (vanilla JS) now expose a state property that reflects the current agent state. Previously, reading state required manually tracking it through the onStateUpdate callback.

React (useAgent)

const agent = useAgent({
	agent: "game-agent",
	name: "room-123",
});

// Read state directly — no separate useState + onStateUpdate needed
return <div>Score: {agent.state?.score}</div>;

// Spread for partial updates
agent.setState({ ...agent.state, score: (agent.state?.score ?? 0) + 10 });

const agent = useAgent<GameAgent, GameState>({
	agent: "game-agent",
	name: "room-123",
});

// Read state directly — no separate useState + onStateUpdate needed
return <div>Score: {agent.state?.score}</div>;

// Spread for partial updates
agent.setState({ ...agent.state, score: (agent.state?.score ?? 0) + 10 });

agent.state is reactive — the component re-renders when state changes from either the server or a client-side setState() call.

Vanilla JS (AgentClient)

const client = new AgentClient({
	agent: "game-agent",
	name: "room-123",
	host: "your-worker.workers.dev",
});

client.setState({ score: 100 });
console.log(client.state); // { score: 100 }

const client = new AgentClient<GameAgent>({
	agent: "game-agent",
	name: "room-123",
	host: "your-worker.workers.dev",
});

client.setState({ score: 100 });
console.log(client.state); // { score: 100 }

State starts as undefined and is populated when the server sends the initial state on connect (from initialState) or when setState() is called. Use optional chaining (agent.state?.field) for safe access. The onStateUpdate callback continues to work as before — the new state property is additive.

schedule() now supports an idempotent option that deduplicates by (type, callback, payload), preventing duplicate rows from accumulating when called in places that run on every Durable Object restart such as onStart().

Cron schedules are idempotent by default. Calling schedule("0 * * * *", "tick") multiple times with the same callback, expression, and payload returns the existing schedule row instead of creating a new one. Pass { idempotent: false } to override.

Delayed and date-scheduled types support opt-in idempotency:

import { Agent } from "agents";

class MyAgent extends Agent {
	async onStart() {
		// Safe across restarts — only one row is created
		await this.schedule(60, "maintenance", undefined, { idempotent: true });
	}
}

import { Agent } from "agents";

class MyAgent extends Agent {
	async onStart() {
		// Safe across restarts — only one row is created
		await this.schedule(60, "maintenance", undefined, { idempotent: true });
	}
}

Two new warnings help catch common foot-guns:

• Calling schedule() inside onStart() without { idempotent: true } emits a console.warn with actionable guidance (once per callback; skipped for cron and when idempotent is set explicitly).
• If an alarm cycle processes 10 or more stale one-shot rows for the same callback, the SDK emits a console.warn and a schedule:duplicate_warning diagnostics channel event.

AgentClient now accepts an optional agent type parameter for full type inference on RPC calls, matching the typed experience already available with useAgent.

const client = new AgentClient({
	agent: "my-agent",
	host: window.location.host,
});

// Typed call — method name autocompletes, args and return type inferred
const value = await client.call("getValue");

// Typed stub — direct RPC-style proxy
await client.stub.getValue();
await client.stub.add(1, 2);

const client = new AgentClient<MyAgent>({
	agent: "my-agent",
	host: window.location.host,
});

// Typed call — method name autocompletes, args and return type inferred
const value = await client.call("getValue");

// Typed stub — direct RPC-style proxy
await client.stub.getValue();
await client.stub.add(1, 2);

State is automatically inferred from the agent type, so onStateUpdate is also typed:

const client = new AgentClient({
	agent: "my-agent",
	host: window.location.host,
	onStateUpdate: (state) => {
		// state is typed as MyAgent's state type
	},
});

const client = new AgentClient<MyAgent>({
	agent: "my-agent",
	host: window.location.host,
	onStateUpdate: (state) => {
		// state is typed as MyAgent's state type
	},
});

Existing untyped usage continues to work without changes. The RPC type utilities (AgentMethods, AgentStub, RPCMethods) are now exported from agents/client for advanced typing scenarios. agents, @cloudflare/ai-chat, and @cloudflare/codemode now require zod ^4.0.0. Zod v3 is no longer supported.

• Turn serialization — onChatMessage() and _reply() work is now queued so user requests, tool continuations, and saveMessages() never stream concurrently.
• Duplicate messages on stop — Clicking stop during an active stream no longer splits the assistant message into two entries.
• Duplicate messages after tool calls — Orphaned client IDs no longer leak into persistent storage.

keepAlive() now uses a lightweight in-memory ref count instead of schedule rows. Multiple concurrent callers share a single alarm cycle. The @experimental tag has been removed from both keepAlive() and keepAliveWhile().

A new entry point @cloudflare/codemode/tanstack-ai adds support for TanStack AI's ↗ chat() as an alternative to the Vercel AI SDK's streamText():

import {
	createCodeTool,
	tanstackTools,
} from "@cloudflare/codemode/tanstack-ai";
import { chat } from "@tanstack/ai";

const codeTool = createCodeTool({
	tools: [tanstackTools(myServerTools)],
	executor,
});

const stream = chat({ adapter, tools: [codeTool], messages });

import { createCodeTool, tanstackTools } from "@cloudflare/codemode/tanstack-ai";
import { chat } from "@tanstack/ai";

const codeTool = createCodeTool({
	tools: [tanstackTools(myServerTools)],
	executor,
});

const stream = chat({ adapter, tools: [codeTool], messages });

To update to the latest version:

npm i agents@latest @cloudflare/ai-chat@latest

The latest releases of @cloudflare/codemode ↗ add a new MCP barrel export, remove ai and zod as required peer dependencies from the main entry point, and give you more control over the sandbox.

A new @cloudflare/codemode/mcp entry point provides two functions that wrap MCP servers with Code Mode:

• codeMcpServer({ server, executor }) — wraps an existing MCP server with a single code tool where each upstream tool becomes a typed codemode.* method.
• openApiMcpServer({ spec, executor, request }) — creates search and execute MCP tools from an OpenAPI spec with host-side request proxying and automatic $ref resolution.

import { codeMcpServer } from "@cloudflare/codemode/mcp";
import { DynamicWorkerExecutor } from "@cloudflare/codemode";

const executor = new DynamicWorkerExecutor({ loader: env.LOADER });

// Wrap an existing MCP server — all its tools become
// typed methods the LLM can call from generated code
const server = await codeMcpServer({ server: upstreamMcp, executor });

import { codeMcpServer } from "@cloudflare/codemode/mcp";
import { DynamicWorkerExecutor } from "@cloudflare/codemode";

const executor = new DynamicWorkerExecutor({ loader: env.LOADER });

// Wrap an existing MCP server — all its tools become
// typed methods the LLM can call from generated code
const server = await codeMcpServer({ server: upstreamMcp, executor });

Breaking change in v0.2.0: generateTypes and the ToolDescriptor / ToolDescriptors types have moved to @cloudflare/codemode/ai:

// Before
import { generateTypes } from "@cloudflare/codemode";

// After
import { generateTypes } from "@cloudflare/codemode/ai";

// Before
import { generateTypes } from "@cloudflare/codemode";

// After
import { generateTypes } from "@cloudflare/codemode/ai";

The main entry point (@cloudflare/codemode) no longer requires the ai or zod peer dependencies. It now exports:

The ai and zod peer dependencies are now optional — only required when importing from @cloudflare/codemode/ai.

DynamicWorkerExecutor now accepts an optional modules option to inject custom ES modules into the sandbox:

const executor = new DynamicWorkerExecutor({
	loader: env.LOADER,
	modules: {
		"utils.js": `export function add(a, b) { return a + b; }`,
	},
});

// Sandbox code can then: import { add } from "utils.js"

const executor = new DynamicWorkerExecutor({
	loader: env.LOADER,
	modules: {
		"utils.js": `export function add(a, b) { return a + b; }`,
	},
});

// Sandbox code can then: import { add } from "utils.js"

DynamicWorkerExecutor now normalizes code and sanitizes tool names internally. You no longer need to call normalizeCode() or sanitizeToolName() before passing code and functions to execute().

npm i @cloudflare/codemode@latest

See the Code Mode documentation for the full API reference.

Sandboxes now support real-time filesystem watching via sandbox.watch(). The method returns a Server-Sent Events ↗ stream backed by native inotify, so your Worker receives create, modify, delete, and move events as they happen inside the container.

Pass a directory path and optional filters. The returned stream is a standard ReadableStream you can proxy directly to a browser client or consume server-side.

// Stream events to a browser client
const stream = await sandbox.watch("/workspace/src", {
	recursive: true,
	include: ["*.ts", "*.js"],
});

return new Response(stream, {
	headers: { "Content-Type": "text/event-stream" },
});

// Stream events to a browser client
const stream = await sandbox.watch("/workspace/src", {
	recursive: true,
	include: ["*.ts", "*.js"],
});

return new Response(stream, {
	headers: { "Content-Type": "text/event-stream" },
});

Use parseSSEStream to iterate over events inside a Worker without forwarding them to a client.

import { parseSSEStream } from "@cloudflare/sandbox";

const stream = await sandbox.watch("/workspace/src", { recursive: true });

for await (const event of parseSSEStream(stream)) {
	console.log(event.type, event.path);
}

import { parseSSEStream } from "@cloudflare/sandbox";
import type { FileWatchSSEEvent } from "@cloudflare/sandbox";

const stream = await sandbox.watch("/workspace/src", { recursive: true });

for await (const event of parseSSEStream<FileWatchSSEEvent>(stream)) {
	console.log(event.type, event.path);
}

Each event includes a type field (create, modify, delete, or move) and the affected path. Move events also include a from field with the original path.

To update to the latest version:

npm i @cloudflare/sandbox@latest

For full API details, refer to the Sandbox file watching reference.

The latest release of the Agents SDK ↗ rewrites observability from scratch with diagnostics_channel, adds keepAlive() to prevent Durable Object eviction during long-running work, and introduces waitForMcpConnections so MCP tools are always available when onChatMessage runs.

The previous observability system used console.log() with a custom Observability.emit() interface. v0.7.0 replaces it with structured events published to diagnostics channels — silent by default, zero overhead when nobody is listening.

Every event has a type, payload, and timestamp. Events are routed to seven named channels:

Use the typed subscribe() helper from agents/observability for type-safe access:

import { subscribe } from "agents/observability";

const unsub = subscribe("rpc", (event) => {
	if (event.type === "rpc") {
		console.log(`RPC call: ${event.payload.method}`);
	}
	if (event.type === "rpc:error") {
		console.error(
			`RPC failed: ${event.payload.method} — ${event.payload.error}`,
		);
	}
});

// Clean up when done
unsub();

import { subscribe } from "agents/observability";

const unsub = subscribe("rpc", (event) => {
	if (event.type === "rpc") {
		console.log(`RPC call: ${event.payload.method}`);
	}
	if (event.type === "rpc:error") {
		console.error(
			`RPC failed: ${event.payload.method} — ${event.payload.error}`,
		);
	}
});

// Clean up when done
unsub();

In production, all diagnostics channel messages are automatically forwarded to Tail Workers — no subscription code needed in the agent itself:

export default {
	async tail(events) {
		for (const event of events) {
			for (const msg of event.diagnosticsChannelEvents) {
				// msg.channel is "agents:rpc", "agents:workflow", etc.
				console.log(msg.timestamp, msg.channel, msg.message);
			}
		}
	},
};

export default {
	async tail(events) {
		for (const event of events) {
			for (const msg of event.diagnosticsChannelEvents) {
				// msg.channel is "agents:rpc", "agents:workflow", etc.
				console.log(msg.timestamp, msg.channel, msg.message);
			}
		}
	},
};

The custom Observability override interface is still supported for users who need to filter or forward events to external services.

For the full event reference, refer to the Observability documentation.

Durable Objects are evicted after a period of inactivity (typically 70-140 seconds with no incoming requests, WebSocket messages, or alarms). During long-running operations — streaming LLM responses, waiting on external APIs, running multi-step computations — the agent can be evicted mid-flight.

keepAlive() prevents this by creating a 30-second heartbeat schedule. The alarm firing resets the inactivity timer. Returns a disposer function that cancels the heartbeat when called.

const dispose = await this.keepAlive();
try {
	const result = await longRunningComputation();
	await sendResults(result);
} finally {
	dispose();
}

const dispose = await this.keepAlive();
try {
	const result = await longRunningComputation();
	await sendResults(result);
} finally {
	dispose();
}

keepAliveWhile() wraps an async function with automatic cleanup — the heartbeat starts before the function runs and stops when it completes:

const result = await this.keepAliveWhile(async () => {
	const data = await longRunningComputation();
	return data;
});

const result = await this.keepAliveWhile(async () => {
	const data = await longRunningComputation();
	return data;
});

Key details:

• Multiple concurrent callers — Each keepAlive() call returns an independent disposer. Disposing one does not affect others.
• AIChatAgent built-in — AIChatAgent automatically calls keepAlive() during streaming responses. You do not need to add it yourself.
• Uses the scheduling system — The heartbeat does not conflict with your own schedules. It shows up in getSchedules() if you need to inspect it.

For the full API reference and when-to-use guidance, refer to Schedule tasks — Keeping the agent alive.

AIChatAgent now waits for MCP server connections to settle before calling onChatMessage. This ensures this.mcp.getAITools() returns the full set of tools, especially after Durable Object hibernation when connections are being restored in the background.

export class ChatAgent extends AIChatAgent {
	// Default — waits up to 10 seconds
	// waitForMcpConnections = { timeout: 10_000 };

	// Wait forever
	waitForMcpConnections = true;

	// Disable waiting
	waitForMcpConnections = false;
}

export class ChatAgent extends AIChatAgent {
	// Default — waits up to 10 seconds
	// waitForMcpConnections = { timeout: 10_000 };

	// Wait forever
	waitForMcpConnections = true;

	// Disable waiting
	waitForMcpConnections = false;
}

For lower-level control, call this.mcp.waitForConnections() directly inside onChatMessage instead.

• MCP deduplication by name and URL — addMcpServer with HTTP transport now deduplicates on both server name and URL. Calling it with the same name but a different URL creates a new connection. URLs are normalized before comparison (trailing slashes, default ports, hostname case).
• callbackHost optional for non-OAuth servers — addMcpServer no longer requires callbackHost when connecting to MCP servers that do not use OAuth.
• MCP URL security — Server URLs are validated before connection to prevent SSRF. Private IP ranges, loopback addresses, link-local addresses, and cloud metadata endpoints are blocked.
• Custom denial messages — addToolOutput now supports state: "output-error" with errorText for custom denial messages in human-in-the-loop tool approval flows.
• requestId in chat options — onChatMessage options now include a requestId for logging and correlating events.

To update to the latest version:

npm i agents@latest @cloudflare/ai-chat@latest

The latest release of the Agents SDK ↗ lets you define an Agent and an McpAgent in the same Worker and connect them over RPC — no HTTP, no network overhead. It also makes OAuth opt-in for simple MCP connections, hardens the schema converter for production workloads, and ships a batch of @cloudflare/ai-chat reliability fixes.

You can now connect an Agent to an McpAgent in the same Worker using a Durable Object binding instead of an HTTP URL. The connection stays entirely within the Cloudflare runtime — no network round-trips, no serialization overhead.

Pass the Durable Object namespace directly to addMcpServer:

import { Agent } from "agents";

export class MyAgent extends Agent {
	async onStart() {
		// Connect via DO binding — no HTTP, no network overhead
		await this.addMcpServer("counter", env.MY_MCP);

		// With props for per-user context
		await this.addMcpServer("counter", env.MY_MCP, {
			props: { userId: "user-123", role: "admin" },
		});
	}
}

import { Agent } from "agents";

export class MyAgent extends Agent {
	async onStart() {
		// Connect via DO binding — no HTTP, no network overhead
		await this.addMcpServer("counter", env.MY_MCP);

		// With props for per-user context
		await this.addMcpServer("counter", env.MY_MCP, {
			props: { userId: "user-123", role: "admin" },
		});
	}
}

The addMcpServer method now accepts string | DurableObjectNamespace as the second parameter with full TypeScript overloads, so HTTP and RPC paths are type-safe and cannot be mixed.

Key capabilities:

• Hibernation support — RPC connections survive Durable Object hibernation automatically. The binding name and props are persisted to storage and restored on wake-up, matching the behavior of HTTP MCP connections.
• Deduplication — Calling addMcpServer with the same server name returns the existing connection instead of creating duplicates. Connection IDs are stable across hibernation restore.
• Smaller surface area — The RPC transport internals have been rewritten and reduced from 609 lines to 245 lines. RPCServerTransport now uses JSONRPCMessageSchema from the MCP SDK for validation instead of hand-written checks.

addMcpServer() no longer eagerly creates an OAuth provider for every connection. For servers that do not require authentication, a simple call is all you need:

// No callbackHost, no OAuth config — just works
await this.addMcpServer("my-server", "https://mcp.example.com");

// No callbackHost, no OAuth config — just works
await this.addMcpServer("my-server", "https://mcp.example.com");

If the server responds with a 401, the SDK throws a clear error: "This MCP server requires OAuth authentication. Provide callbackHost in addMcpServer options to enable the OAuth flow." The restore-from-storage flow also handles missing callback URLs gracefully, skipping auth provider creation for non-OAuth servers.

The schema converter used by generateTypes() and getAITools() now handles edge cases that previously caused crashes in production:

• Depth and circular reference guards — Prevents stack overflows on recursive or deeply nested schemas
• $ref resolution — Supports internal JSON Pointers (#/definitions/..., #/$defs/..., #)
• Tuple support — prefixItems (JSON Schema 2020-12) and array items (draft-07)
• OpenAPI 3.0 nullable: true — Supported across all schema branches
• Per-tool error isolation — One malformed schema cannot crash the full pipeline in generateTypes() or getAITools()
• Missing inputSchema fallback — getAITools() falls back to { type: "object" } instead of throwing

• Tool denial flow — Denied tool approvals (approved: false) now transition to output-denied with a tool_result, fixing Anthropic provider compatibility. Custom denial messages are supported via state: "output-error" and errorText.
• Abort/cancel support — Streaming responses now properly cancel the reader loop when the abort signal fires and send a done signal to the client.
• Duplicate message persistence — persistMessages() now reconciles assistant messages by content and order, preventing duplicate rows when clients resend full history.
• requestId in OnChatMessageOptions — Handlers can now send properly-tagged error responses for pre-stream failures.
• redacted_thinking preservation — The message sanitizer no longer strips Anthropic redacted_thinking blocks.
• /get-messages reliability — Endpoint handling moved from a prototype onRequest() override to a constructor wrapper, so it works even when users override onRequest without calling super.onRequest().
• Client tool APIs undeprecated — createToolsFromClientSchemas, clientTools, AITool, extractClientToolSchemas, and the tools option on useAgentChat are restored for SDK use cases where tools are defined dynamically at runtime.
• jsonSchema initialization — Fixed jsonSchema not initialized error when calling getAITools() in onChatMessage.

To update to the latest version:

npm i agents@latest @cloudflare/ai-chat@latest

Sandboxes now support createBackup() and restoreBackup() methods for creating and restoring point-in-time snapshots of directories.

This allows you to restore environments quickly. For instance, in order to develop in a sandbox, you may need to include a user's codebase and run a build step. Unfortunately git clone and npm install can take minutes, and you don't want to run these steps every time the user starts their sandbox.

Now, after the initial setup, you can just call createBackup(), then restoreBackup() the next time this environment is needed. This makes it practical to pick up exactly where a user left off, even after days of inactivity, without repeating expensive setup steps.

const sandbox = getSandbox(env.Sandbox, "my-sandbox");

// Make non-trivial changes to the file system
await sandbox.gitCheckout(endUserRepo, { targetDir: "/workspace" });
await sandbox.exec("npm install", { cwd: "/workspace" });

// Create a point-in-time backup of the directory
const backup = await sandbox.createBackup({ dir: "/workspace" });

// Store the handle for later use
await env.KV.put(`backup:${userId}`, JSON.stringify(backup));

// ... in a future session...

// Restore instead of re-cloning and reinstalling
await sandbox.restoreBackup(backup);

Backups are stored in R2 and can take advantage of R2 object lifecycle rules to ensure they do not persist forever.

Key capabilities:

• Persist and reuse across sandbox sessions — Easily store backup handles in KV, D1, or Durable Object storage for use in subsequent sessions
• Usable across multiple instances — Fork a backup across many sandboxes for parallel work
• Named backups — Provide optional human-readable labels for easier management
• TTLs — Set time-to-live durations so backups are automatically removed from storage once they are no longer needed

To get started, refer to the backup and restore guide for setup instructions and usage patterns, or the Backups API reference for full method documentation.

The @cloudflare/codemode ↗ package has been rewritten into a modular, runtime-agnostic SDK.

Code Mode ↗ enables LLMs to write and execute code that orchestrates your tools, instead of calling them one at a time. This can (and does) yield significant token savings, reduces context window pressure and improves overall model performance on a task.

The new Executor interface is runtime agnostic and comes with a prebuilt DynamicWorkerExecutor to run generated code in a Dynamic Worker Loader.

• Removed experimental_codemode() and CodeModeProxy — the package no longer owns an LLM call or model choice
• New import path: createCodeTool() is now exported from @cloudflare/codemode/ai

• createCodeTool() — Returns a standard AI SDK Tool to use in your AI agents.
• Executor interface — Minimal execute(code, fns) contract. Implement for any code sandboxing primitive or runtime.

Runs code in a Dynamic Worker. It comes with the following features:

• Network isolation — fetch() and connect() blocked by default (globalOutbound: null) when using DynamicWorkerExecutor
• Console capture — console.log/warn/error captured and returned in ExecuteResult.logs
• Execution timeout — Configurable via timeout option (default 30s)

import { createCodeTool } from "@cloudflare/codemode/ai";
import { DynamicWorkerExecutor } from "@cloudflare/codemode";
import { streamText } from "ai";

const executor = new DynamicWorkerExecutor({ loader: env.LOADER });
const codemode = createCodeTool({ tools: myTools, executor });

const result = streamText({
	model,
	tools: { codemode },
	messages,
});

import { createCodeTool } from "@cloudflare/codemode/ai";
import { DynamicWorkerExecutor } from "@cloudflare/codemode";
import { streamText } from "ai";

const executor = new DynamicWorkerExecutor({ loader: env.LOADER });
const codemode = createCodeTool({ tools: myTools, executor });

const result = streamText({
	model,
	tools: { codemode },
	messages,
});

{
	"worker_loaders": [{ "binding": "LOADER" }],
}

[[worker_loaders]]
binding = "LOADER"

See the Code Mode documentation for full API reference and examples.

npm i @cloudflare/codemode@latest

The latest release of the Agents SDK ↗ adds built-in retry utilities, per-connection protocol message control, and a fully rewritten @cloudflare/ai-chat with data parts, tool approval persistence, and zero breaking changes.

A new this.retry() method lets you retry any async operation with exponential backoff and jitter. You can pass an optional shouldRetry predicate to bail early on non-retryable errors.

class MyAgent extends Agent {
	async onRequest(request) {
		const data = await this.retry(() => callUnreliableService(), {
			maxAttempts: 4,
			shouldRetry: (err) => !(err instanceof PermanentError),
		});
		return Response.json(data);
	}
}

class MyAgent extends Agent {
	async onRequest(request: Request) {
		const data = await this.retry(() => callUnreliableService(), {
			maxAttempts: 4,
			shouldRetry: (err) => !(err instanceof PermanentError),
		});
		return Response.json(data);
	}
}

Retry options are also available per-task on queue(), schedule(), scheduleEvery(), and addMcpServer():

// Per-task retry configuration, persisted in SQLite alongside the task
await this.schedule(
	Date.now() + 60_000,
	"sendReport",
	{ userId: "abc" },
	{
		retry: { maxAttempts: 5 },
	},
);

// Class-level retry defaults
class MyAgent extends Agent {
	static options = {
		retry: { maxAttempts: 3 },
	};
}

// Per-task retry configuration, persisted in SQLite alongside the task
await this.schedule(Date.now() + 60_000, "sendReport", { userId: "abc" }, {
	retry: { maxAttempts: 5 },
});

// Class-level retry defaults
class MyAgent extends Agent {
	static options = {
		retry: { maxAttempts: 3 },
	};
}

Retry options are validated eagerly at enqueue/schedule time, and invalid values throw immediately. Internal retries have also been added for workflow operations (terminateWorkflow, pauseWorkflow, and others) with Durable Object-aware error detection.

Agents automatically send JSON text frames (identity, state, MCP server lists) to every WebSocket connection. You can now suppress these per-connection for clients that cannot handle them — binary-only devices, MQTT clients, or lightweight embedded systems.

class MyAgent extends Agent {
	shouldSendProtocolMessages(connection, ctx) {
		// Suppress protocol messages for MQTT clients
		const subprotocol = ctx.request.headers.get("Sec-WebSocket-Protocol");
		return subprotocol !== "mqtt";
	}
}

class MyAgent extends Agent {
	shouldSendProtocolMessages(connection: Connection, ctx: ConnectionContext) {
		// Suppress protocol messages for MQTT clients
		const subprotocol = ctx.request.headers.get("Sec-WebSocket-Protocol");
		return subprotocol !== "mqtt";
	}
}

Connections with protocol messages disabled still fully participate in RPC and regular messaging. Use isConnectionProtocolEnabled(connection) to check a connection's status at any time. The flag persists across Durable Object hibernation.

See Protocol messages for full documentation.

The first stable release of @cloudflare/ai-chat ships alongside this release with a major refactor of AIChatAgent internals — new ResumableStream class, WebSocket ChatTransport, and simplified SSE parsing — with zero breaking changes. Existing code using AIChatAgent and useAgentChat works as-is.

Key new features:

• Data parts — Attach typed JSON blobs (data-*) to messages alongside text. Supports reconciliation (type+id updates in-place), append, and transient parts (ephemeral via onData callback). See Data parts.
• Tool approval persistence — The needsApproval approval UI now survives page refresh and DO hibernation. The streaming message is persisted to SQLite when a tool enters approval-requested state.
• maxPersistedMessages — Cap SQLite message storage with automatic oldest-message deletion.
• body option on useAgentChat — Send custom data with every request (static or dynamic).
• Incremental persistence — Hash-based cache to skip redundant SQL writes.
• Row size guard — Automatic two-pass compaction when messages approach the SQLite 2 MB limit.
• autoContinueAfterToolResult defaults to true — Client-side tool results and tool approvals now automatically trigger a server continuation, matching server-executed tool behavior. Set autoContinueAfterToolResult: false in useAgentChat to restore the previous behavior.

Notable bug fixes:

• Resolved stream resumption race conditions
• Resolved an issue where setMessages functional updater sent empty arrays
• Resolved an issue where client tool schemas were lost after DO hibernation
• Resolved InvalidPromptError after tool approval (approval.id was dropped)
• Resolved an issue where message metadata was not propagated on broadcast/resume paths
• Resolved an issue where clearAll() did not clear in-memory chunk buffers
• Resolved an issue where reasoning-delta silently dropped data when reasoning-start was missed during stream resumption

getQueue(), getQueues(), getSchedule(), dequeue(), dequeueAll(), and dequeueAllByCallback() were unnecessarily async despite only performing synchronous SQL operations. They now return values directly instead of wrapping them in Promises. This is backward compatible — existing code using await on these methods will continue to work.

• Fix TypeScript "excessively deep" error — A depth counter on CanSerialize and IsSerializableParam types bails out to true after 10 levels of recursion, preventing the "Type instantiation is excessively deep" error with deeply nested types like AI SDK CoreMessage[].
• POST SSE keepalive — The POST SSE handler now sends event: ping every 30 seconds to keep the connection alive, matching the existing GET SSE handler behavior. This prevents POST response streams from being silently dropped by proxies during long-running tool calls.
• Widened peer dependency ranges — Peer dependency ranges across packages have been widened to prevent cascading major bumps during 0.x minor releases. @cloudflare/ai-chat and @cloudflare/codemode are now marked as optional peer dependencies.

To update to the latest version:

npm i agents@latest @cloudflare/ai-chat@latest

We're excited to announce GLM-4.7-Flash on Workers AI, a fast and efficient text generation model optimized for multilingual dialogue and instruction-following tasks, along with the brand-new @cloudflare/tanstack-ai ↗ package and workers-ai-provider v3.1.1 ↗.

You can now run AI agents entirely on Cloudflare. With GLM-4.7-Flash's multi-turn tool calling support, plus full compatibility with TanStack AI and the Vercel AI SDK, you have everything you need to build agentic applications that run completely at the edge.

@cf/zai-org/glm-4.7-flash is a multilingual model with a 131,072 token context window, making it ideal for long-form content generation, complex reasoning tasks, and multilingual applications.

Key Features and Use Cases:

• Multi-turn Tool Calling for Agents: Build AI agents that can call functions and tools across multiple conversation turns
• Multilingual Support: Built to handle content generation in multiple languages effectively
• Large Context Window: 131,072 tokens for long-form writing, complex reasoning, and processing long documents
• Fast Inference: Optimized for low-latency responses in chatbots and virtual assistants
• Instruction Following: Excellent at following complex instructions for code generation and structured tasks

Use GLM-4.7-Flash through the Workers AI binding (env.AI.run()), the REST API at /run or /v1/chat/completions, AI Gateway, or via workers-ai-provider for the Vercel AI SDK.

Pricing is available on the model page or pricing page.

We've released @cloudflare/tanstack-ai, a new package that brings Workers AI and AI Gateway support to TanStack AI ↗. This provides a framework-agnostic alternative for developers who prefer TanStack's approach to building AI applications.

Workers AI adapters support four configuration modes — plain binding (env.AI), plain REST, AI Gateway binding (env.AI.gateway(id)), and AI Gateway REST — across all capabilities:

• Chat (createWorkersAiChat) — Streaming chat completions with tool calling, structured output, and reasoning text streaming.
• Image generation (createWorkersAiImage) — Text-to-image models.
• Transcription (createWorkersAiTranscription) — Speech-to-text.
• Text-to-speech (createWorkersAiTts) — Audio generation.
• Summarization (createWorkersAiSummarize) — Text summarization.

AI Gateway adapters route requests from third-party providers — OpenAI, Anthropic, Gemini, Grok, and OpenRouter — through Cloudflare AI Gateway for caching, rate limiting, and unified billing.

To get started:

npm install @cloudflare/tanstack-ai @tanstack/ai

The Workers AI provider for the Vercel AI SDK ↗ now supports three new capabilities beyond chat and image generation:

• Transcription (provider.transcription(model)) — Speech-to-text with automatic handling of model-specific input formats across binding and REST paths.
• Text-to-speech (provider.speech(model)) — Audio generation with support for voice and speed options.
• Reranking (provider.reranking(model)) — Document reranking for RAG pipelines and search result ordering.

import { createWorkersAI } from "workers-ai-provider";
import {
	experimental_transcribe,
	experimental_generateSpeech,
	rerank,
} from "ai";

const workersai = createWorkersAI({ binding: env.AI });

const transcript = await experimental_transcribe({
	model: workersai.transcription("@cf/openai/whisper-large-v3-turbo"),
	audio: audioData,
	mediaType: "audio/wav",
});

const speech = await experimental_generateSpeech({
	model: workersai.speech("@cf/deepgram/aura-1"),
	text: "Hello world",
	voice: "asteria",
});

const ranked = await rerank({
	model: workersai.reranking("@cf/baai/bge-reranker-base"),
	query: "What is machine learning?",
	documents: ["ML is a branch of AI.", "The weather is sunny."],
});

This release also includes a comprehensive reliability overhaul (v3.0.5):

• Fixed streaming — Responses now stream token-by-token instead of buffering all chunks, using a proper TransformStream pipeline with backpressure.
• Fixed tool calling — Resolved issues with tool call ID sanitization, conversation history preservation, and a heuristic that silently fell back to non-streaming mode when tools were defined.
• Premature stream termination detection — Streams that end unexpectedly now report finishReason: "error" instead of silently reporting "stop".
• AI Search support — Added createAISearch as the canonical export (renamed from AutoRAG). createAutoRAG still works with a deprecation warning.

To upgrade:

npm install workers-ai-provider@latest ai

• @cloudflare/tanstack-ai on npm ↗
• workers-ai-provider on npm ↗
• GitHub repository ↗

The latest release of the Agents SDK ↗ brings readonly connections, MCP protocol and security improvements, x402 payment protocol v2 migration, and the ability to customize OAuth for MCP server connections.

Agents can now restrict WebSocket clients to read-only access, preventing them from modifying agent state. This is useful for dashboards, spectator views, or any scenario where clients should observe but not mutate.

New hooks: shouldConnectionBeReadonly, setConnectionReadonly, isConnectionReadonly. Readonly connections block both client-side setState() and mutating @callable() methods, and the readonly flag survives hibernation.

class MyAgent extends Agent {
	shouldConnectionBeReadonly(connection) {
		// Make spectators readonly
		return connection.url.includes("spectator");
	}
}

class MyAgent extends Agent {
	shouldConnectionBeReadonly(connection) {
		// Make spectators readonly
		return connection.url.includes("spectator");
	}
}

The new createMcpOAuthProvider method on the Agent class allows subclasses to override the default OAuth provider used when connecting to MCP servers. This enables custom authentication strategies such as pre-registered client credentials or mTLS, beyond the built-in dynamic client registration.

class MyAgent extends Agent {
	createMcpOAuthProvider(callbackUrl) {
		return new MyCustomOAuthProvider(this.ctx.storage, this.name, callbackUrl);
	}
}

class MyAgent extends Agent {
	createMcpOAuthProvider(callbackUrl: string): AgentMcpOAuthProvider {
		return new MyCustomOAuthProvider(this.ctx.storage, this.name, callbackUrl);
	}
}

Upgraded the MCP SDK to 1.26.0 to prevent cross-client response leakage. Stateless MCP Servers should now create a new McpServer instance per request instead of sharing a single instance. A guard is added in this version of the MCP SDK which will prevent connection to a Server instance that has already been connected to a transport. Developers will need to modify their code if they declare their McpServer instance as a global variable.

Added callbackPath option to addMcpServer to prevent instance name leakage in MCP OAuth callback URLs. When sendIdentityOnConnect is false, callbackPath is now required — the default callback URL would expose the instance name, undermining the security intent. Also fixes callback request detection to match via the state parameter instead of a loose /callback URL substring check, enabling custom callback paths.

onStateChanged is a drop-in rename of onStateUpdate (same signature, same behavior). onStateUpdate still works but emits a one-time console warning per class. validateStateChange rejections now propagate a CF_AGENT_STATE_ERROR message back to the client.

Migrated the x402 MCP payment integration from the legacy x402 package to @x402/core and @x402/evm v2.

Breaking changes for x402 users:

• Peer dependencies changed: replace x402 with @x402/core and @x402/evm
• PaymentRequirements type now uses v2 fields (e.g. amount instead of maxAmountRequired)
• X402ClientConfig.account type changed from viem.Account to ClientEvmSigner (structurally compatible with privateKeyToAccount())

npm uninstall x402
npm install @x402/core @x402/evm

Network identifiers now accept both legacy names and CAIP-2 format:

// Legacy name (auto-converted)
{
	network: "base-sepolia",
}

// CAIP-2 format (preferred)
{
	network: "eip155:84532",
}

Other x402 changes:

• X402ClientConfig.network is now optional — the client auto-selects from available payment requirements
• Server-side lazy initialization: facilitator connection is deferred until the first paid tool invocation
• Payment tokens support both v2 (PAYMENT-SIGNATURE) and v1 (X-PAYMENT) HTTP headers
• Added normalizeNetwork export for converting legacy network names to CAIP-2 format
• Re-exports PaymentRequirements, PaymentRequired, Network, FacilitatorConfig, and ClientEvmSigner from agents/x402

• Fix useAgent and AgentClient crashing when using basePath routing
• CORS handling delegated to partyserver's native support (simpler, more reliable)
• Client-side onStateUpdateError callback for handling rejected state updates

To update to the latest version:

npm i agents@latest

The Sandbox SDK ↗ now supports PTY (pseudo-terminal) passthrough, enabling browser-based terminal UIs to connect to sandbox shells via WebSocket.

The new terminal() method proxies a WebSocket upgrade to the container's PTY endpoint, with output buffering for replay on reconnect.

// Worker: proxy WebSocket to container terminal
return sandbox.terminal(request, { cols: 80, rows: 24 });

// Worker: proxy WebSocket to container terminal
return sandbox.terminal(request, { cols: 80, rows: 24 });

Each session can have its own terminal with an isolated working directory and environment, so users can run separate shells side-by-side in the same container.

// Multiple isolated terminals in the same sandbox
const dev = await sandbox.getSession("dev");
return dev.terminal(request);

// Multiple isolated terminals in the same sandbox
const dev = await sandbox.getSession("dev");
return dev.terminal(request);

The new @cloudflare/sandbox/xterm export provides a SandboxAddon for xterm.js ↗ with automatic reconnection (exponential backoff + jitter), buffered output replay, and resize forwarding.

import { SandboxAddon } from "@cloudflare/sandbox/xterm";

const addon = new SandboxAddon({
	getWebSocketUrl: ({ sandboxId, origin }) =>
		`${origin}/ws/terminal?id=${sandboxId}`,
	onStateChange: (state, error) => updateUI(state),
});
terminal.loadAddon(addon);
addon.connect({ sandboxId: "my-sandbox" });

import { SandboxAddon } from "@cloudflare/sandbox/xterm";

const addon = new SandboxAddon({
	getWebSocketUrl: ({ sandboxId, origin }) =>
		`${origin}/ws/terminal?id=${sandboxId}`,
	onStateChange: (state, error) => updateUI(state),
});
terminal.loadAddon(addon);
addon.connect({ sandboxId: "my-sandbox" });

To update to the latest version:

npm i @cloudflare/sandbox@latest

The latest release of the Agents SDK ↗ brings first-class support for Cloudflare Workflows, synchronous state management, and new scheduling capabilities.

Agents excel at real-time communication and state management. Workflows excel at durable execution. Together, they enable powerful patterns where Agents handle WebSocket connections while Workflows handle long-running tasks, retries, and human-in-the-loop flows.

Use the new AgentWorkflow class to define workflows with typed access to your Agent:

import { AgentWorkflow } from "agents/workflows";

export class ProcessingWorkflow extends AgentWorkflow {
	async run(event, step) {
		// Call Agent methods via RPC
		await this.agent.updateStatus(event.payload.taskId, "processing");

		// Non-durable: progress reporting to clients
		await this.reportProgress({ step: "process", percent: 0.5 });
		this.broadcastToClients({ type: "update", taskId: event.payload.taskId });

		// Durable via step: idempotent, won't repeat on retry
		await step.mergeAgentState({ taskProgress: 0.5 });

		const result = await step.do("process", async () => {
			return processData(event.payload.data);
		});

		await step.reportComplete(result);
		return result;
	}
}

import { AgentWorkflow } from "agents/workflows";
import type { AgentWorkflowEvent, AgentWorkflowStep } from "agents/workflows";

export class ProcessingWorkflow extends AgentWorkflow<MyAgent, TaskParams> {
	async run(event: AgentWorkflowEvent<TaskParams>, step: AgentWorkflowStep) {
		// Call Agent methods via RPC
		await this.agent.updateStatus(event.payload.taskId, "processing");

		// Non-durable: progress reporting to clients
		await this.reportProgress({ step: "process", percent: 0.5 });
		this.broadcastToClients({ type: "update", taskId: event.payload.taskId });

		// Durable via step: idempotent, won't repeat on retry
		await step.mergeAgentState({ taskProgress: 0.5 });

		const result = await step.do("process", async () => {
			return processData(event.payload.data);
		});

		await step.reportComplete(result);
		return result;
	}
}

Start workflows from your Agent with runWorkflow() and handle lifecycle events:

export class MyAgent extends Agent {
	async startTask(taskId, data) {
		const instanceId = await this.runWorkflow("PROCESSING_WORKFLOW", {
			taskId,
			data,
		});
		return { instanceId };
	}

	async onWorkflowProgress(workflowName, instanceId, progress) {
		this.broadcast(JSON.stringify({ type: "progress", progress }));
	}

	async onWorkflowComplete(workflowName, instanceId, result) {
		console.log(`Workflow ${instanceId} completed`);
	}

	async onWorkflowError(workflowName, instanceId, error) {
		console.error(`Workflow ${instanceId} failed:`, error);
	}
}

export class MyAgent extends Agent {
	async startTask(taskId: string, data: string) {
		const instanceId = await this.runWorkflow("PROCESSING_WORKFLOW", {
			taskId,
			data,
		});
		return { instanceId };
	}

	async onWorkflowProgress(
		workflowName: string,
		instanceId: string,
		progress: unknown,
	) {
		this.broadcast(JSON.stringify({ type: "progress", progress }));
	}

	async onWorkflowComplete(
		workflowName: string,
		instanceId: string,
		result?: unknown,
	) {
		console.log(`Workflow ${instanceId} completed`);
	}

	async onWorkflowError(
		workflowName: string,
		instanceId: string,
		error: unknown,
	) {
		console.error(`Workflow ${instanceId} failed:`, error);
	}
}

Key workflow methods on your Agent:

• runWorkflow(workflowName, params, options?) — Start a workflow with optional metadata
• getWorkflow(workflowId) / getWorkflows(criteria?) — Query workflows with cursor-based pagination
• approveWorkflow(workflowId) / rejectWorkflow(workflowId) — Human-in-the-loop approval flows
• pauseWorkflow(), resumeWorkflow(), terminateWorkflow() — Workflow control

State updates are now synchronous with a new validateStateChange() validation hook:

export class MyAgent extends Agent {
	validateStateChange(oldState, newState) {
		// Return false to reject the change
		if (newState.count < 0) return false;
		// Return modified state to transform
		return { ...newState, lastUpdated: Date.now() };
	}
}

export class MyAgent extends Agent<Env, State> {
	validateStateChange(oldState: State, newState: State): State | false {
		// Return false to reject the change
		if (newState.count < 0) return false;
		// Return modified state to transform
		return { ...newState, lastUpdated: Date.now() };
	}
}