agents 0.27.0 → 0.28.0

package/docs/fetch-the-web.md

@@ -0,0 +1,276 @@
+# Fetch the Web (Beta)
+
+`agents/webfetch` gives a model a `web_fetch` tool that reads one URL and returns the most useful text form of it: Markdown for HTML pages, PDFs, and Office documents, pretty-printed JSON, or plain text as served. Long content is windowed, so the model reads a page in parts. The same tool is available for the pi harness, the AI SDK, and TanStack AI.
+
+The host decides which URLs can be fetched, how much is read, and how long a fetch may take. The model only chooses the URL, where to continue reading, and whether to get the body converted or raw.
+
+> **Beta** — this feature may have breaking changes in future releases.
+
+## Quick Start
+
+You need an `AI` binding in `wrangler.jsonc`. The tool fetches pages with the Worker's own `fetch` and uses `env.AI.toMarkdown()` from [Workers AI Markdown Conversion](https://developers.cloudflare.com/workers-ai/features/markdown-conversion/) to turn HTML and documents into Markdown. There is no API token to set up.
+
+```jsonc
+{
+  // No local implementation: under `wrangler dev` or `vite dev`, conversions
+  // run against the account you are logged in to.
+  "ai": { "binding": "AI", "remote": true }
+}
+```
+
+Pages that are already Markdown, JSON, or plain text never reach the binding. The tool asks for `text/markdown` first, so sites with [Markdown for Agents](https://developers.cloudflare.com/fundamentals/reference/markdown-for-agents/) enabled answer in Markdown and skip conversion.
+
+Then add the tool to your harness. Every adapter takes the same options; the TanStack AI adapter also takes `name`, because TanStack AI tools carry their name in the definition.
+
+Pi harness:
+
+```ts
+import { webFetchTool } from "agents/webfetch/pi";
+
+this.registry.install({
+  name: "tools",
+  tools: [webFetchTool({ binding: this.env.AI })]
+});
+```
+
+AI SDK:
+
+```ts
+import { webFetchTool } from "agents/webfetch/ai-sdk";
+
+const result = streamText({
+  model,
+  tools: { web_fetch: webFetchTool({ binding: this.env.AI }) },
+  messages
+});
+```
+
+If you convert UI messages yourself, pass the same tools to `convertToModelMessages(messages, { tools })`. Without them, the AI SDK sends earlier fetches back to the model as the JSON output: the same window, but without the header line or the `<untrusted_web_content>` wrapper. `AiSdkHarness` does this for you.
+
+TanStack AI:
+
+```ts
+import { webFetchTool } from "agents/webfetch/tanstack-ai";
+
+const tools = [webFetchTool({ binding: this.env.AI })];
+```
+
+## Options
+
+An invalid value throws a `RangeError` when the tool is created.
+
+| Option              | Default                 | Notes                                                                                                                                         |
+| ------------------- | ----------------------- | --------------------------------------------------------------------------------------------------------------------------------------------- |
+| `binding`           | —                       | The `AI` binding, for converting HTML and documents. Or pass `source` instead (see [Other sources](#other-sources)).                          |
+| `pageChars`         | `20000`                 | Characters the model reads per call. It continues with `offset`.                                                                              |
+| `maxBytes`          | `5000000`               | Stop reading a response body past this many bytes, as a `web_fetch_too_large` failure. Bodies are streamed, so a large one is never buffered. |
+| `timeoutMs`         | `30000`                 | Give up on a fetch after this long, redirects and conversion included, as a retryable `web_fetch_timeout` failure.                            |
+| `maxRedirects`      | `5`                     | Redirects to follow, from 0 to 20.                                                                                                            |
+| `allowedHosts`      | any public host         | Host patterns URLs must match: `example.com`, `*.example.com` (subdomains, not the apex), or `*`.                                             |
+| `blockedHosts`      | —                       | Host patterns never fetched, in the same syntax. Checked first.                                                                               |
+| `allowPrivateHosts` | `false`                 | Allow loopback, private, and local-only hosts. For local development against `localhost`; see [URL policy](#url-policy).                      |
+| `userAgent`         | `Cloudflare-Agents (…)` | The `User-Agent` header.                                                                                                                      |
+| `description`       | built-in description    | Replaces the tool description the model sees.                                                                                                 |
+
+With `source`, only `pageChars`, `timeoutMs`, and `description` apply: the URL policy and limits belong to the source.
+
+## Model Interface
+
+The model sends:
+
+| Field    | Notes                                                                                                                |
+| -------- | -------------------------------------------------------------------------------------------------------------------- |
+| `url`    | The http(s) URL to read, at most 2,000 characters.                                                                   |
+| `offset` | Character offset to continue a long page from, as the previous result says. Default `0`.                             |
+| `format` | `"auto"` (default) converts to the most readable form; `"raw"` returns a textual body exactly as the server sent it. |
+
+The model gets text: a header line, then one window of the content inside `<untrusted_web_content>`:
+
+```
+web_fetch: https://developers.cloudflare.com/agents/ · 200 · text/html → markdown · chars 0–19873 of 81200 · continue with offset=19873
+<untrusted_web_content>
+# Agents
+…
+</untrusted_web_content>
+```
+
+The header line names the final URL after redirects, the status, the content type and how it was read, and where the window sits. A window ends on a line break when one is close to its end, so `offset` values are rarely round numbers; the model continues from the offset the header gives. When the window reaches the end, the header has no `continue with`. An error status reads `404 Not Found` rather than `404`, so the model cannot miss it.
+
+The wrapper and the tool description tell the model to treat the content as data, not instructions. A page cannot close the wrapper early: a closing tag inside the content is escaped.
+
+`renderWebFetchPage(output)` from `agents/webfetch` is the renderer, if you want the same text elsewhere. It formats a stored output as is, so re-rendering an old result gives the text the model read.
+
+## Host Output
+
+The host gets `WebFetchToolOutput`: the window the model read and the page's metadata. The rest of the content is not stored, so a long page does not bloat the conversation history:
+
+| Field                   | Notes                                                                                                          |
+| ----------------------- | -------------------------------------------------------------------------------------------------------------- |
+| `url`, `finalUrl`       | The URL as requested, and after redirects.                                                                     |
+| `status`                | The final response's HTTP status.                                                                              |
+| `contentType`           | The media type, lowercased, without parameters.                                                                |
+| `via`                   | How the content was produced; see [Conversion](#conversion).                                                   |
+| `content`, `totalChars` | The window of converted content the model read, and the whole page's length.                                   |
+| `offset`, `nextOffset`  | The window the model read, and where the next one starts (`null` at the end).                                  |
+| `title`                 | The page title, from `<title>` for HTML or the first `# ` heading for Markdown, when there is one.             |
+| `headers`               | `content-type`, `content-length`, `last-modified`, `etag`, `x-markdown-tokens`, and `cf-mitigated`, when sent. |
+| `bytes`                 | Body bytes read.                                                                                               |
+| `redirects`             | The URLs that redirected, in order; `finalUrl` is not included.                                                |
+
+Where it arrives depends on the harness:
+
+- **Pi**: in the tool result's `details`, as `{ ok: true, output }`. The tool is `replay: "safe"`: if a fetch is interrupted mid-call, for example by an eviction, pi runs it again when the session recovers. Completed results are stored and not fetched again.
+- **AI SDK**: as the return value of `execute`, so `onFinish`, UI message parts, and logs see it. `toModelOutput` renders it for the model.
+- **TanStack AI**: the server tool returns the rendered window, so the host gets the same text the model does.
+
+Each call fetches the page again; there is no cache between calls, so reading a long page in five windows fetches it five times.
+
+## Conversion
+
+With `format: "auto"`, the content type decides what the model reads:
+
+| Content type                                                                                                         | Handling                                               | `via`                 |
+| -------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------ | --------------------- |
+| `text/markdown`                                                                                                      | Passed through                                         | `markdown-negotiated` |
+| `text/html`, `application/xhtml+xml`                                                                                 | Converted with `env.AI.toMarkdown()`                   | `converted`           |
+| `application/pdf`, Word (`.docx`), Excel (`.xlsx`, `.xlsm`, `.xlsb`, `.xls`), OpenDocument (`.odt`, `.ods`), Numbers | Converted with `env.AI.toMarkdown()`                   | `converted`           |
+| `application/json`, `*+json`                                                                                         | Pretty-printed; invalid JSON is passed through as text | `json`                |
+| Other `text/*`, `application/xml`, `*+xml`, JavaScript, YAML, TOML, NDJSON, SQL, GraphQL, shell                      | Passed through                                         | `text`                |
+| Images, audio, video, fonts, archives, binary `application/octet-stream`, anything else                              | Rejected as `web_fetch_unsupported_content_type`       | —                     |
+
+Bodies are decoded with the charset from a byte-order mark, the `Content-Type` header, or an HTML `<meta charset>`, falling back to UTF-8. A response without a `Content-Type`, or sent as `application/octet-stream` (common for `.md`, `.yaml`, and source files), is sniffed as HTML, PDF, binary, or plain text.
+
+With `format: "raw"`, textual bodies, HTML included, are returned exactly as served with `via: "raw"`. Documents and binary bodies are rejected, because their raw bytes are not text.
+
+A 4xx or 5xx response with a readable body is returned like any page, with its status in the header line: error pages often explain what went wrong. Only an empty or non-textual error body becomes a `web_fetch_http_error` failure. A bot-management challenge page (sent with `cf-mitigated: challenge`) is reported as `web_fetch_blocked` rather than returned as content, because a plain fetch cannot pass it; the model is told to use a browser tool or another source.
+
+## URL Policy
+
+Every URL is checked before it is fetched, and every redirect target is checked again before it is followed. A redirect from `https:` down to `http:` is refused. Requests are GET only, with no cookies, credentials, or host-supplied headers.
+
+Always rejected:
+
+- Schemes other than `http:` and `https:`.
+- URLs with a username or password.
+- URLs longer than 2,000 characters.
+
+Rejected unless `allowPrivateHosts` is `true`:
+
+- Loopback, private, CGNAT, link-local (including cloud metadata at `169.254.169.254`), multicast, and reserved addresses, in IPv4 or IPv6, including IPv4-mapped, NAT64, and 6to4 forms. The list is a subset of the IANA special-purpose address registries ([RFC 6890](https://www.rfc-editor.org/rfc/rfc6890)).
+- `localhost` and names ending in `.localhost` or `.internal`.
+
+Hosts are checked after the URL parser has canonicalised them, so a private address cannot hide behind an unusual spelling such as `2130706433`, `0x7f.1`, `0177.0.0.1`, or `[::ffff:127.0.0.1]`. Names are not resolved: a public name that points at a private address, such as `127.0.0.1.nip.io`, passes the policy. In production, Workers egress cannot reach private ranges anyway; the policy is for local development, other runtimes, and defence in depth. The fragment is dropped from every URL.
+
+`blockedHosts` is checked first, then `allowedHosts`, then the private-host rule. To keep the model on your documentation:
+
+```ts
+webFetchTool({
+  binding: this.env.AI,
+  allowedHosts: ["developers.cloudflare.com", "*.cloudflare.com"]
+});
+```
+
+No DNS lookup is done: a public name that resolves to a private address is the runtime's to refuse. Deployed Workers cannot reach private networks. Under `wrangler dev` they can, which is why the private-host rule exists, and why `allowPrivateHosts` is the switch for testing against a server on `localhost`:
+
+```ts
+webFetchTool({
+  binding: this.env.AI,
+  // Local development only: lets the model reach http://localhost:8787.
+  allowPrivateHosts: true
+});
+```
+
+## Failures
+
+A failed fetch becomes a `WebFetchError` with `status`, `code`, `retryable`, and, where there is one, `url`, `contentType`, and (for a timeout) `timeoutMs`. Every error has a `code`:
+
+| Code                                 | Retryable              | When                                                                     |
+| ------------------------------------ | ---------------------- | ------------------------------------------------------------------------ |
+| `invalid_web_fetch_input`            | No                     | The URL, `offset`, or `format` is malformed, or the URL is too long.     |
+| `web_fetch_disallowed_url`           | No                     | The URL's scheme, credentials, or host is not allowed by the URL policy. |
+| `web_fetch_disallowed_redirect`      | No                     | A redirect pointed at a URL the policy does not allow, or https to http. |
+| `web_fetch_too_many_redirects`       | No                     | More than `maxRedirects` redirects, or a redirect loop.                  |
+| `web_fetch_timeout`                  | Yes                    | The fetch took longer than `timeoutMs`.                                  |
+| `web_fetch_too_large`                | No                     | The body is larger than `maxBytes`.                                      |
+| `web_fetch_unsupported_content_type` | No                     | The body is binary, or a document requested with `format: "raw"`.        |
+| `web_fetch_http_error`               | 5xx, 408, and 429 only | A 4xx or 5xx response with an empty or non-textual body.                 |
+| `web_fetch_blocked`                  | No                     | The site answered with a bot challenge (`cf-mitigated: challenge`).      |
+| `web_fetch_conversion_failed`        | Yes                    | Workers AI could not convert the body to Markdown.                       |
+| `web_fetch_unavailable`              | Yes                    | The network failed: DNS, TLS, the connection, or reading the body.       |
+
+The error's `message` is written for you. The model gets different text that tells it what to do next: fix the input, retry once, retry with `format: "raw"` after a failed HTML conversion (a document that fails to convert is not retried), or move on.
+
+- **Pi**: the tool returns an error result with the model's text, and `details` is `{ ok: false, message, status, code, retryable, url?, contentType? }`.
+- **AI SDK** and **TanStack AI**: the tool throws a `WebFetchError`, which is how those frameworks report tool errors. Its `message` is the model's text, and its `cause` is the original error with the detail.
+
+Every adapter entry point re-exports `WebFetchError`, so `instanceof` checks do not need a second import.
+
+If the harness cancels the call, the fetch is aborted and the abort propagates instead of becoming a failed result.
+
+## With Web Search
+
+`web_fetch` pairs with [`web_search`](./search-the-web.md): search finds pages, fetch reads them. Each tool's description names the other, so a model with both searches, then fetches the results it wants to read.
+
+```ts
+import { webFetchTool } from "agents/webfetch/ai-sdk";
+import { webSearchTool } from "agents/websearch/ai-sdk";
+
+const tools = {
+  web_search: webSearchTool({ binding: this.env.AI }),
+  web_fetch: webFetchTool({ binding: this.env.AI })
+};
+```
+
+## What It Is Not
+
+- **Not an HTTP client.** There is no POST, no request body, and no custom headers, cookies, or credentials. For calls to your own APIs, write a tool for them.
+- **Not a browser.** Pages are not rendered, so content that only appears after JavaScript runs is missing. For rendered pages, screenshots, or interaction, see [Browse the Web](./browse-the-web.md).
+- **Not a summariser.** The model reads the page itself, a window at a time.
+
+## Other Sources
+
+`fetchWeb` from `agents/webfetch` fetches one URL with no model involved: the same pipeline, URL policy, limits, and `timeoutMs`. It returns the whole page as a `WebFetchPage`: the fields of `WebFetchToolOutput` with all of `content`, and no `offset` or `nextOffset`. Use it from scheduled jobs, callables, or tests:
+
+```ts
+import { WebFetchError, fetchWeb } from "agents/webfetch";
+
+try {
+  const page = await fetchWeb(
+    { url: "https://developers.cloudflare.com/agents/" },
+    { binding: this.env.AI, maxBytes: 1_000_000 }
+  );
+  console.log(page.title, page.via, page.totalChars);
+} catch (error) {
+  if (error instanceof WebFetchError) console.log(error.code, error.retryable);
+}
+```
+
+To show a page a window at a time without a model, for example in a UI that pages through what the model would read, cut it with `windowWebFetchPage` and render each window with `renderWebFetchPage`. This is exactly what the tool does, and it does not fetch again:
+
+```ts
+import { renderWebFetchPage, windowWebFetchPage } from "agents/webfetch";
+
+const first = windowWebFetchPage(page, { offset: 0, pageChars: 20_000 });
+console.log(renderWebFetchPage(first));
+// Next window: windowWebFetchPage(page, { offset: first.nextOffset, … }),
+// until nextOffset is null.
+```
+
+`createDirectWebFetch` from `agents/webfetch` returns a `WebFetchSource`: an object whose `fetch({ url, format? }, { signal? })` returns a `WebFetchPage`. A source validates `url` and `format` itself (the tool checks only `offset`), throwing `invalid_web_fetch_input` for a malformed request. Any `WebFetchSource` can be passed to a tool as `source` instead of `binding`. That is how tests substitute a fake, and how you wrap a source with caching or logging:
+
+```ts
+import { createDirectWebFetch } from "agents/webfetch";
+import { webFetchTool } from "agents/webfetch/ai-sdk";
+
+const direct = createDirectWebFetch({ binding: this.env.AI });
+
+webFetchTool({
+  source: {
+    async fetch(request, options) {
+      const page = await direct.fetch(request, options);
+      console.log(request.url, page.status, page.bytes);
+      return page;
+    }
+  }
+});
+```

package/docs/harnesses/pi.md

@@ -125,6 +125,31 @@ harness: async ({ storage, context }) => {
 
 The existing `@cloudflare/think` skills work out-of-the-box.
 
+### Browser
+
+`browserTool` from `agents/browser/pi` gives the model a persistent browser it drives with the Chrome DevTools Protocol. Put a `Browser` on the object's Lifecycle and install the tool in an extension:
+
+```ts
+import { Browser, browserRun } from "agents/browser";
+import { browserTool } from "agents/browser/pi";
+
+readonly browser = new Browser({ provider: browserRun(this.env.BROWSER) });
+
+harness: async ({ storage, context }) => {
+  this.registry.install({
+    name: "browser",
+    tools: [
+      browserTool({ ctx: this.ctx, browser: this.browser, loader: this.env.LOADER })
+    ]
+  });
+  return Harness.open(storage, { models, registry: this.registry }, context);
+};
+
+readonly lifecycle = Lifecycle.install(this).use(this.browser).use(this.harness);
+```
+
+It needs a `browser` binding, a `LOADER` Worker Loader binding, and `export { CodemodeRuntime } from "@cloudflare/codemode"` from the Worker entry. Screenshots come back as images, and each conversation works in its own tab. Refer to [Persistent browser](../browse-the-web.md#persistent-browser) for how it behaves.
+
 ## Work with sessions
 
 A session is a Pi conversation. Each has its own transcript, inbox, model and run, and they can run at the same time.

package/docs/index.md

@@ -1,6 +1,25 @@
 # Agents Documentation
 
-Build stateful AI agents on Cloudflare Workers. Every agent is a Durable Object — an addressable, hibernatable actor with its own SQLite database, WebSockets, and scheduling — so you can afford one durable agent per user, account, task, or conversation, with near-zero cost while idle.
+Build stateful AI agents on Cloudflare Workers. An agent is a Durable Object: an addressable actor with its own SQLite database, alarms, and WebSockets. You can run one durable agent per user, account, task, or conversation, and pay close to nothing while it is idle.
+
+Every agent starts the same way. Extend `DurableObject`, install a [Lifecycle](./lifecycle.md), and add the capabilities you need:
+
+```ts
+import { DurableObject } from "cloudflare:workers";
+import { Lifecycle } from "agents/lifecycle";
+import { MCPClientManager } from "agents/mcp/client";
+import { Scheduler } from "agents/schedules";
+
+export class Assistant extends DurableObject<Env> {
+  readonly mcp = new MCPClientManager("assistant", "1.0.0");
+  readonly scheduler = new Scheduler({ callbacks: {} });
+  readonly lifecycle = Lifecycle.install(this)
+    .use(this.mcp)
+    .use(this.scheduler);
+}
+```
+
+The Lifecycle runs startup, routes requests to capabilities, and owns the object's durable job queue and its single alarm. Capabilities are independent pieces that share it.
 
 ## Related package documentation
 
@@ -9,49 +28,76 @@ Build stateful AI agents on Cloudflare Workers. Every agent is a Durable Object
 
 ## Choose your path
 
-Pick the base class that matches what you are building. They share the same Durable Object foundation, so you can start small and move up without re-platforming.
+| You are building...                                     | Use                                                                                             |
+| ------------------------------------------------------- | ----------------------------------------------------------------------------------------------- |
+| An agent that runs a model loop with tools and recovery | [Pi harness](./harnesses/pi.md) on a Durable Object                                             |
+| Your own agent loop                                     | [Sessions](./sessions.md), [Streams](./streams.md), and [Tasks](./tasks.md) on a Durable Object |
+| A chat or reasoning agent with every default built in   | [`Think`](https://github.com/cloudflare/agents/blob/main/docs/think/index.md)                   |
+| A chat UI that owns the loop and the stream             | [`AIChatAgent`](./chat-agents.md), built on the `Agent` class                                   |
+| Real-time state sync, `@callable` methods, email, voice | The [`Agent` class](./agent-class.md), which installs a preset of capabilities for you          |
+| Durable multi-step processes (not chat)                 | [Workflows](./workflows.md)                                                                     |
 
-| You are building...                                            | Use                                                                           | Why                                                                                                                    |
-| -------------------------------------------------------------- | ----------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------- |
-| Stateful backend logic, real-time sync, custom protocols       | [`Agent`](./agent-class.md)                                                   | The core class: state, WebSockets, scheduling, SQL, and sub-agents. No opinions about chat or LLMs.                    |
-| A chat UI where you own the loop, the stream, and the response | [`AIChatAgent`](./chat-agents.md)                                             | A thin chat-protocol adapter for `useAgentChat`. Bring your own agentic loop and custom streaming.                     |
-| A durable, general-purpose reasoning agent                     | [`Think`](https://github.com/cloudflare/agents/blob/main/docs/think/index.md) | Opinionated runtime: agentic loop, sessions, tools, memory, compaction, recovery, and multi-channel delivery built in. |
-| A voice agent (speech in, speech out)                          | [Voice mixins](./voice.md)                                                    | `withVoice` adds real-time STT/TTS, interruption and barge-in, and conversation persistence to an agent.               |
-| Durable multi-step processes (not chat)                        | [Workflows](./workflows.md)                                                   | Long-running, retryable step orchestration with Cloudflare Workflows.                                                  |
+## Getting started
 
-Not sure? Start with [`Agent`](./agent-class.md) for raw building blocks, or [`Think`](https://github.com/cloudflare/agents/blob/main/docs/think/index.md) if you want a chat or reasoning agent that already handles the hard parts.
+- [Getting Started](./getting-started.md) - Quick start guide for new users
+- [Adding to an Existing Project](./adding-to-existing-project.md) - Integrate agents into your app
 
-## What makes these production-grade
+## Lifecycle
 
-The differentiator is not "we have durable state" — it is what happens when a turn is interrupted. Agents built on this SDK keep their promises across Durable Object eviction, deploys, client disconnects, and human waits:
+- [Durable Object Lifecycle](./lifecycle.md) - Install a Lifecycle on a Durable Object, write capabilities, and use the job queue
+- [getCurrentAgent()](./get-current-agent.md) - Accessing the current object, request, and connection across async calls
 
-- **Turn recovery** — an in-flight LLM turn survives Durable Object eviction and resumes instead of silently dying. See [Chat & Fiber Recovery](./chat-agents.md#stream-recovery) and [Durable Execution](./durable-execution.md).
-- **Resumable streams** — a disconnected client rejoins the same stream rather than losing the response. See [Resumable Streaming](./resumable-streaming.md).
-- **Recovery-aware delivery** — Think snapshots channel delivery as `accepted`, `streaming`, or `completed`, so a restart replays a not-yet-streamed answer but posts a safe interruption notice rather than risking a duplicate partial reply. See [Messengers — Delivery and Recovery](https://github.com/cloudflare/agents/blob/main/docs/think/messengers.md#delivery-and-recovery).
-- **Durable submissions** — webhooks and RPC callers submit a turn with an idempotency key and check status later, instead of holding a request open. See [Programmatic Submissions](https://github.com/cloudflare/agents/blob/main/docs/think/programmatic-submissions.md).
-- **Human-in-the-loop without hangs** — a turn can pause for approval and resume later. A human wait is a first-class state, not a stuck request. See [Human in the Loop](./human-in-the-loop.md).
+## Harnesses
 
-## Getting Started
+- [Pi harness (Beta)](./harnesses/pi.md) - Host pi-durable sessions in a Durable Object with durable storage and lifecycle wakeups
+- [Think harness (Experimental)](./harnesses/think.md) - Run Think's agent loop as a Lifecycle capability, with transcripts in Sessions, output in Streams, and turns that survive eviction
+- [Think (Experimental)](https://github.com/cloudflare/agents/blob/main/docs/think/index.md) - Opinionated chat agent with built-in memory, tools, and streaming. Extends `Think`, which builds on `Agent`.
 
-- [Getting Started](./getting-started.md) - Quick start guide for new users
-- [Adding to an Existing Project](./adding-to-existing-project.md) - Integrate agents into your app
-- [Understanding the Agent Class](./agent-class.md) - Deep dive into the Agent class architecture
+## Capabilities
 
-## Core Concepts
+Install each capability with `lifecycle.use()`. Capabilities marked experimental may change between releases.
 
-- [State Management](./state.md) - Managing agent state with `setState()`, `initialState`, and `onStateChanged()`
+| Capability         | Import              | Status       | Docs                                                            |
+| ------------------ | ------------------- | ------------ | --------------------------------------------------------------- |
+| `PiHarness`        | `agents/harness/pi` | Beta         | [Pi harness](./harnesses/pi.md)                                 |
+| `MCPClientManager` | `agents/mcp/client` | Stable       | [MCP client](./mcp-client.md)                                   |
+| `Scheduler`        | `agents/schedules`  | Stable       | [Scheduling](./scheduling.md)                                   |
+| `State`            | `agents/state`      | Stable       | [Lifecycle](./lifecycle.md#websockets-are-an-opt-in-capability) |
+| `WebSockets`       | `agents/websockets` | Stable       | [Lifecycle](./lifecycle.md#websockets-are-an-opt-in-capability) |
+| `Queue`            | `agents/queue`      | Experimental | [Queue](./queue.md)                                             |
+| `RoutedAgents`     | `agents/routing`    | Experimental | [Routing](./routing.md)                                         |
+| `Browser`          | `agents/browser`    | Experimental | [Browse the Web](./browse-the-web.md)                           |
+| `Sessions`         | `agents/sessions`   | Experimental | [Sessions](./sessions.md)                                       |
+| `Streams`          | `agents/streams`    | Experimental | [Streams](./streams.md)                                         |
+| `Tasks`            | `agents/tasks`      | Experimental | [Tasks](./tasks.md)                                             |
+
+Building your own harness? Sessions stores the conversation, Streams holds durable incremental output, and Tasks runs replayable background work. [Models](./models.md) and [Models for pi-ai](./models-pi-ai.md) give you one provider for Workers AI and third-party models. [Context](./context.md) and [Workspace](https://github.com/cloudflare/agents/blob/main/docs/shell/index.md) are also available.
+
+## Agent class
+
+`Agent` extends `DurableObject`, installs a Lifecycle, and adds `Scheduler`, `Queue`, `MCPClientManager`, `State`, `WebSockets`, `Tasks`, and dynamic agents for you. Its methods (`this.schedule()`, `this.queue()`, `this.setState()`) delegate to those capabilities. Features below are documented for `Agent` and its subclasses.
+
+### Core
+
+- [Understanding the Agent Class](./agent-class.md) - How the Agent class is built on the Lifecycle
+- [State Management](./state.md) - `setState()`, `initialState`, and `onStateChanged()`
 - [Routing](./routing.md) - How `routeAgentRequest()` and agent naming works, plus `RoutedAgents` for a hub that routes to many independent Agents
 - [Dynamic agents](./sub-agents.md) - Facet-backed child agents for code the parent supervises (dynamic/generated code, per-run tool agents, sandboxes) — not the recommended primitive for many independent peers like chats
 - [HTTP & WebSockets](./http-websockets.md) - Request handling and real-time connections
 - [Callable Methods](./callable-methods.md) - The `@callable` decorator and client-server method calls
 - [Readonly Connections](./readonly-connections.md) - Restricting which connections can modify state
-- [getCurrentAgent()](./get-current-agent.md) - Accessing agent context across async calls
+- [Client SDK](./client-sdk.md) - Connecting from React (`useAgent`) and vanilla JS (`AgentClient`), state sync, and RPC calls
 
-## Client SDK
+### Background processing
 
-- [Client SDK](./client-sdk.md) - Connecting from React (`useAgent`) and vanilla JS (`AgentClient`), state sync, and RPC calls
+- [Queue](./queue.md) - Durable background task execution
+- [Scheduling](./scheduling.md) - Delayed, scheduled, and cron-based tasks
+- [Retries](./retries.md) - Automatic retries with exponential backoff and jitter
+- [Durable Execution](./durable-execution.md) - `runFiber()`, `startFiber()`, `stash()`, and crash recovery for long tasks
+- [Workflows](./workflows.md) - Durable multi-step processing with Cloudflare Workflows
+- [Human in the Loop](./human-in-the-loop.md) - Approval flows and manual intervention
 
-## Communication Channels
+### Communication channels
 
 - [Email Service](./email.md) - Sending, receiving, and replying to emails
 - [Webhooks](./webhooks.md) - Receiving and sending webhook events
@@ -60,35 +106,17 @@ The differentiator is not "we have durable state" — it is what happens when a
 - [Voice Agents](./voice.md) - Build voice agents with real-time speech-to-text, text-to-speech, and conversation persistence
 - [Chat SDK State](./chat-sdk.md) - Store Chat SDK subscriptions, locks, queues, and history in Agents sub-agents
 
-## Background Processing
-
-- [Queue](./queue.md) - Durable background task execution
-- [Scheduling](./scheduling.md) - Delayed, scheduled, and cron-based tasks
-- [Retries](./retries.md) - Automatic retries with exponential backoff and jitter
-- [Durable Execution](./durable-execution.md) - `runFiber()`, `startFiber()`, `stash()`, and crash recovery for long tasks
-- [Workflows](./workflows.md) - Durable multi-step processing with Cloudflare Workflows
-- [Human in the Loop](./human-in-the-loop.md) - Approval flows and manual intervention
+### Chat and AI
 
-## AI Integration
-
-- TODO: [AI SDK Integration](./ai-sdk.md) - Using Vercel AI SDK with agents
-- TODO: [TanStack Integration](./tanstack.md) - Using TanStack AI with agents
 - [Chat Agents](./chat-agents.md) - `AIChatAgent` class and `useAgentChat` React hook
 - [Chat & Fiber Recovery](./chat-agents.md#stream-recovery) - Recover LLM turns after Durable Object eviction
 - [Agent Tools](./agent-tools.md) - Run chat-capable sub-agents as tools with streaming child timelines
 - [Server-Driven Messages](./server-driven-messages.md) - Autonomous agent workflows: scheduled follow-ups, queue processing, webhooks, chained reasoning
-- TODO: [Using AI Models](./using-ai-models.md) - OpenAI, Anthropic, Workers AI, and other providers
-- [Models (Experimental)](./models.md) - `createAI` — one AI SDK provider for Workers AI and third-party catalog models, same string space
-- [Models for pi-ai (Beta)](./models-pi-ai.md) - `createAI` for pi-ai: Workers AI ids and third-party models through AI Gateway, as a pi-ai provider
-- [Pi harness (Beta)](./harnesses/pi.md) - Host pi-durable sessions in a Durable Object with durable storage and lifecycle wakeups
-- [Think harness (Experimental)](./harnesses/think.md) - Run Think's agent loop as a Lifecycle capability, with transcripts in Sessions, output in Streams, and turns that survive eviction
-- TODO: [RAG (Retrieval Augmented Generation)](./rag.md) - Vector search with Vectorize
-- [Sessions (Experimental)](./sessions.md) - Durable message trees, streamed history, compaction, search, and lossless attachment offload
-- [Context (Experimental)](./context.md) - System-prompt blocks, frozen prompts, writable/searchable/loadable providers, and their tools
-- [Workspace (Experimental)](https://github.com/cloudflare/agents/blob/main/docs/shell/index.md) - Durable virtual filesystem backed by SQLite + R2
-- [Codemode (Experimental)](https://github.com/cloudflare/agents/blob/main/docs/agents/codemode.md) - LLM-generated executable code for tool orchestration
 - [Client Tools Continuation](./client-tools-continuation.md) - Handling tool calls across client/server
 - [Resumable Streaming](./resumable-streaming.md) - Automatic stream resumption on client disconnect
+- [Long-Running Agents](./long-running-agents.md) - Building agents that persist for weeks or months: lifecycle, recovery, async operations, and planning
+- TODO: [SQL API](./sql.md) - Using `this.sql` for direct database queries
+- TODO: [Memory & Persistence](./memory.md) - Long-term storage patterns
 
 ## Think (Experimental)
 
@@ -103,51 +131,46 @@ The differentiator is not "we have durable state" — it is what happens when a
 - [Sub-agents and Programmatic Turns](https://github.com/cloudflare/agents/blob/main/docs/think/sub-agents.md) - RPC streaming, `saveMessages`, recovery
 - [Programmatic Submissions](https://github.com/cloudflare/agents/blob/main/docs/think/programmatic-submissions.md) - Durable Think turn admission for webhooks and RPC callers
 
+## Models and tools
+
+- TODO: [AI SDK Integration](./ai-sdk.md) - Using Vercel AI SDK with agents
+- TODO: [TanStack Integration](./tanstack.md) - Using TanStack AI with agents
+- TODO: [Using AI Models](./using-ai-models.md) - OpenAI, Anthropic, Workers AI, and other providers
+- [Models (Experimental)](./models.md) - `createAI` — one AI SDK provider for Workers AI and third-party catalog models, same string space
+- [Models for pi-ai (Beta)](./models-pi-ai.md) - `createAI` for pi-ai: Workers AI ids and third-party models through AI Gateway, as a pi-ai provider
+- [Context (Experimental)](./context.md) - System-prompt blocks, frozen prompts, writable/searchable/loadable providers, and their tools
+- [Workspace (Experimental)](https://github.com/cloudflare/agents/blob/main/docs/shell/index.md) - Durable virtual filesystem backed by SQLite + R2
+- [Codemode (Experimental)](./codemode.md) - LLM-generated executable code for tool orchestration
+- [Browse the Web (Experimental)](./browse-the-web.md) - Full CDP access for web inspection, scraping, and debugging
+- [Search the Web (Beta)](./search-the-web.md) - A `websearch` tool over Cloudflare's Web Search API for the pi harness, the AI SDK, and TanStack AI
+- [Fetch the Web (Beta)](./fetch-the-web.md) - A `web_fetch` tool that reads a URL as Markdown, JSON, or text, with a host-controlled URL policy, for the pi harness, the AI SDK, TanStack AI, and Think
+- TODO: [Cloudflare Sandboxes](./sandboxes.md) - Isolated environments for coding agents, ffmpeg, and heavy compute
+- TODO: [RAG (Retrieval Augmented Generation)](./rag.md) - Vector search with Vectorize
+
 ## MCP (Model Context Protocol)
 
 - [Creating MCP Servers](./mcp-servers.md) - Build MCP servers with `McpAgent`
 - [Securing MCP Servers](./securing-mcp-servers.md) - OAuth and authentication for MCP
-- [Connecting to MCP Servers](./mcp-client.md) - `addMcpServer()` and consuming external MCP tools
+- [Connecting to MCP Servers](./mcp-client.md) - `MCPClientManager` on a Durable Object, or `addMcpServer()` on an Agent
 - [MCP Transports](./mcp-transports.md) - Transport options: Streamable HTTP, SSE, and RPC
 
-## Authentication & Security
-
-- TODO: [Securing your Agents](./securing-agents.md) - Authentication, authorization, and access control
-- [Cross-Domain Authentication](./cross-domain-authentication.md) - Auth across different domains
-
-## Observability & Debugging
+## Operations
 
+- [Configuration](./configuration.md) - wrangler.jsonc setup, types, secrets, and deployment
 - [Observability](./observability.md) - Monitoring and tracing agent activity
+- [Cross-Domain Authentication](./cross-domain-authentication.md) - Auth across different domains
+- TODO: [Securing your Agents](./securing-agents.md) - Authentication, authorization, and access control
 - TODO: [Testing](./testing.md) - Unit tests, integration tests, mocking agents
 - TODO: [Evals](./evals.md) - Evaluating AI agent quality and behavior
-
-## Agent Studio
-
 - TODO: [Agent Studio](./agent-studio.md) - Local dev tool for inspecting and interacting with agent instances
 
-## Compute Environments
-
-- [Browse the Web (Experimental)](./browse-the-web.md) - Full CDP access for web inspection, scraping, and debugging
-- [Search the Web (Beta)](./search-the-web.md) - A `websearch` tool over Cloudflare's Web Search API for the pi harness, the AI SDK, and TanStack AI
-- TODO: [Cloudflare Sandboxes](./sandboxes.md) - Isolated environments for coding agents, ffmpeg, and heavy compute
-
-## Advanced Topics
-
-- [Long-Running Agents](./long-running-agents.md) - Building agents that persist for weeks or months: lifecycle, recovery, async operations, and planning
-- TODO: [SQL API](./sql.md) - Using `this.sql` for direct database queries
-- TODO: [Memory & Persistence](./memory.md) - Long-term storage patterns
-- [Configuration](./configuration.md) - wrangler.jsonc setup, types, secrets, and deployment
-
-## Migration Guides
+## Migration guides
 
 - [Migration to AI SDK v5](./migration-to-ai-sdk-v5.md)
 - [Migration to AI SDK v6](./migration-to-ai-sdk-v6.md)
 
 ## Reference
 
-- [Durable Object Lifecycle](./lifecycle.md) - Compose reusable durable components outside the Agent base class
-- [Tasks](./tasks.md) - Durable, replayable background work with journaled steps and durable sleeps (experimental)
-- [Streams](./streams.md) - Durable incremental output: chunk log, cursor, replay-then-tail reads (experimental)
 - TODO: [API Reference](./api-reference.md) - Complete API documentation
 - TODO: [FAQ / How is this different from Durable Objects?](./faq.md)
 - TODO: [Resources & Further Reading](./resources.md)

package/docs/lifecycle.md

@@ -1,8 +1,7 @@
 # Durable Object lifecycle
 
-> **Experimental.** Everything exported from `agents/lifecycle` — and the
-> capabilities built on it, including `Scheduler` — may change between
-> releases while the composition surface stabilizes.
+`agents/lifecycle` is stable. Most capabilities built on it are still
+experimental; each capability's page says which.
 
 `agents/lifecycle` lets reusable durable capabilities work in both `Agent` and a
 plain Cloudflare Durable Object. It uses composition: your class extends the

package/docs/queue.md

@@ -7,9 +7,8 @@ survives the object leaving memory.
 
 ## Queue Lifecycle capability
 
-> **Experimental.** The `Queue` capability and the `agents/lifecycle` surface
-> it builds on may change between releases. Agent's established queue methods
-> (`this.queue()` and friends) are stable.
+> **Experimental.** The `Queue` capability may change between releases.
+> Agent's queue methods (`this.queue()` and friends) are stable.
 
 `Queue` is a reusable Lifecycle capability. A plain Lifecycle Object can
 install it without extending `Agent`:

package/docs/scheduling.md

@@ -17,10 +17,6 @@ Under the hood, scheduling uses [Durable Object alarms](https://developers.cloud
 
 ## Scheduler Lifecycle primitive
 
-> **Experimental.** The `Scheduler` primitive and the `agents/lifecycle`
-> surface it builds on may change between releases. Agent's established
-> scheduling methods (`this.schedule()` and friends) are stable.
-
 `Scheduler` is a reusable Lifecycle capability. A plain Lifecycle Object can
 install it without extending `Agent`:
 
@@ -836,8 +832,8 @@ When using this schema with OpenAI models via the AI SDK, you must pass `provide
 
 ## API Reference
 
-Two surfaces share these semantics: the experimental `Scheduler` primitive
-(`agents/schedules`, methods on the scheduler instance) and the stable `Agent`
+Two surfaces share these semantics: the `Scheduler` capability
+(`agents/schedules`, methods on the scheduler instance) and the `Agent`
 methods (on the Agent class, delegating to `this.scheduler`).
 
 ### Scheduler primitive

package/docs/search-the-web.md

@@ -88,6 +88,8 @@ The host gets the full API response, without trimming, as `WebSearchToolOutput`:
 
 The API has no pagination. The tool description tells the model to search again with a rephrased query when it wants more or different results.
 
+Results are for discovery. To let the model read the pages it finds, add the [`web_fetch` tool](./fetch-the-web.md) beside `web_search`; the `web_search` description tells the model to fetch result URLs with `web_fetch` when it is available.
+
 ## Failures
 
 A failed search becomes a `WebSearchError` with `status`, `code`, `retryable`, and `requestId` (AI Gateway's id for the request, for the gateway log). Every error has a `code`: the API's own when it sends one, for example `web_search_payment_required`, otherwise one derived from the HTTP status, such as `web_search_rate_limited` or `web_search_unavailable`. On a runtime older than the one above, the binding has no `websearch()` and the search fails with code `web_search_unsupported_runtime`.