Hey, Seth here.

Issue #4 of Agent Patch Notes, the weekly email from Ranked Agents, where we rank AI agents and automation tools from S to B tier. This week agents moved closer to your own browser and desktop: Browser Use rebuilt itself around Chrome DevTools, OpenCode let agents drive your real browser, and n8n Labs shipped a personal agent that runs on your computer. Also, the MCP reference servers patched six security holes, so it’s a good week to update.

Biggest releases this week

MCP reference servers 2.4.0 fix six security vulnerabilities in Filesystem, Git and Fetch (Nov 11)

Filesystem 2.3.0 reads files relative to the directory it checked, so a directory swapped for a symlink can't leak outside files on Linux and macOS (only narrowed on Windows). It also refuses FIFOs, sockets and devices, and fixes a move_file escape. mcp-server-git stops an untrusted repo's own .git/config (fsmonitor, diff drivers, hooks, filters, signing) from running programs and refuses Git older than 2.31. mcp-server-fetch now treats IPv6 site-local addresses as private.

Browser Use 1.0 rebuilds the agent as a Python REPL on Chrome DevTools Protocol (Nov 9)

A ground-up rewrite that deletes about 80% of the library. The agent now writes Python in one persistent REPL that drives Chrome directly over CDP, with no DOM serializer or action schema. Browser Use reports up to 33 points higher benchmark scores on the same models at 2.7 to 5.2x lower cost and 2 to 3x the speed (self-reported). Defaults change: Browser() is now a Browser Use Cloud browser (Browser('chromium') for local), one BROWSER_USE_API_KEY replaces per-provider keys, and max_steps drops to 100. Check MIGRATING.md before you upgrade.

Claude Code 2.1.319 to 2.1.322: Claude Fable 5.2 default, background MCP tasks, deny-instead-of-prompt (Nov 6–9)

Claude Fable 5.2 is now the default Fable model, with 1M context at $8/$40 per Mtok. When an MCP server runs a tool call as a task, Claude Code moves it to the background and reports back when it's done. The new permissions.denyInsteadOfPrompting setting denies tool calls that would wait on a permission prompt, so the turn keeps going. 2.1.321 also fixes MCP OAuth sending refresh tokens to another host on redirect and a case where a read-only Bash command ran outside the sandbox without a prompt.

n8n Labs releases Nod, an open-source personal agent that runs on your computer (Nov 10)

A research preview, separate from n8n and needing no n8n account. It’s a desktop agent for Mac and Windows (or your own server) with its own local sandbox VM and browser, working in apps like Gmail, Calendar, Drive and Slack. A separate permission check confirms before it sends, pays, posts or deletes, and every action and its cost are logged. Apache 2.0 and free; you bring your own model provider, with a monthly spend limit.

OpenCode launches Browse so agents can drive your real browser (Nov 7)

An MCP server plus Chromium extension that lets OpenCode, Claude Code, Cursor, Codex, Zed or any MCP client use your real browser with your tabs, logins and extensions. Agents work in background tab groups, take screenshots and capture HAR traffic, and hand the tab back to you for logins, 2FA or payments. Per-session site allowlists limit which sites agents can touch.

Perplexity launches Teammates, agents that own ongoing work (Nov 4)

Personal agents that plan and coordinate long-running projects inside Perplexity Computer, remember goals, task history and files across runs, watch connected tools and act when things change. They can hand work to each other and join Slack channels. Research preview on the web for individual Max subscribers, with iOS coming soon.

OpenAI launches Businesses in ChatGPT with Business Agents (Nov 10)

In beta, businesses can claim their profile in ChatGPT through Business Hub and see insights like mentions and website clicks. Eligible U.S. ecommerce and online-service businesses can launch a Business Agent that answers shoppers' questions in a clearly labeled conversation, separate from ChatGPT's own answers. Shopify merchants can connect from Business Hub or the Shopify admin.

There's more on the daily releases feed, including Pydantic AI 2.67 durable SubAgents, CrewAI 1.15.36 to 1.15.38 (gpt-6-luna is now the default model), Mastra 1.82 Knowledge v2, LangGraph 1.2.20 strict resume, Kiro bots GA, Render Sandboxes in public beta, goose 1.58 and 1.59, OpenHands 1.37 and 1.38, and Vercel Labs agent-desktop 1.2.0 with MCP.

Tier watch

No tier changes this week. Nod, OpenCode Browse, Business Agents, Kiro bots and Render Sandboxes aren’t ranked on the site yet.

Still worth watching: Browser Use stays in A tier after 1.0. We updated its profile for the new REPL-on-CDP design and will see how the rewrite holds up before moving it.

For reference, the current S tier is: Claude Code, Cursor, OpenAI Codex, n8n, LangGraph, ChatGPT Deep Research, Zapier, ChatGPT Agent and Vapi. Full tier list

Head-to-head of the week: Browser Use vs Browserbase

Browser Use just shipped 1.0, so here’s how it stacks up against Browserbase:

Browser Use wins if you want an AI agent that does the browsing for you, or the cheapest cloud browsers: $0.02 per browser hour, pay as you go, no subscription, and credits that never expire.

Browserbase wins if you're building your own automation on Playwright, Puppeteer, Selenium or Stagehand and want managed browser infrastructure: monthly plans with included hours (Developer $20/mo, Startup $99/mo), captcha solving on paid plans, stealth and session recordings.

Both ship an official hosted MCP server. Browserbase's lets your assistant drive the browser click by click; Browser Use's hands off a whole task and returns the result.

One tip

Update your MCP reference servers this week, and pin the filesystem one. The 2.4.0 release fixes six vulnerabilities, including an untrusted repo's .git/config running programs through mcp-server-git. If you're on a date-versioned filesystem package, install @modelcontextprotocol/[email protected] explicitly, because you won't get the fix otherwise. Any agent that can read a repo or a folder you didn't write yourself is the one to patch first.

That’s issue #4. Hit reply and tell me which tool you want ranked next. I read every reply, and your answers shape what gets added next.

If a friend or teammate is trying to pick an AI agent, forward this to them. They can subscribe at rankedagents.dev.

See you next Friday,

Seth at Ranked Agents

Keep Reading