The Best AI Agents

OpenAI has reorganized its agent features this year. Its help center now says the older ChatGPT agent mode “is no longer available” and points users to ChatGPT Work for longer, multi-step tasks. According to OpenAI, ChatGPT is now split into Chat, for quick questions, and Work, “an agent designed for longer, multi-step work and finished deliverables.” Work can research a topic, analyze information, and create documents, spreadsheets, presentations and reports. It runs in the cloud on the web and in the mobile apps, and in the desktop app it can also use local files and desktop apps with your permission, where your plan allows it. OpenAI’s help center says Plus includes its GPT-6 Astra model in Work and Codex.

For web tasks, Work uses what OpenAI calls the cloud browser: “its own browser on a separate computer in the cloud.” OpenAI’s examples include checking restaurant availability, comparing flights, tracking a package from details in a connected email account, preparing a DMV appointment for your approval, and reconciling invoices in accounting software. You don’t pick the browser yourself; ChatGPT decides when to use it and may use a connected app instead when that’s more direct. Because it runs remotely, a task can continue after you close your laptop. It pauses when it needs input, a sign-in or a confirmation.

The safety defaults are the most thoughtful we found. By default, ChatGPT asks permission before opening a new website, and you can switch to auto-approve (it pauses only for sites that look unsafe) or allow everything, which OpenAI says it doesn’t recommend. Separately, it’s designed to ask for confirmation before actions that are hard to reverse or create a financial, legal or account commitment, such as confirming a booking or making a payment. When a site needs a login, you type your credentials into a secure form that goes straight to the remote browser; OpenAI says the model can’t see them and ChatGPT doesn’t store them, and a separate review model checks sign-in requests for signs of phishing. The cloud browser keeps its own cookies and doesn’t touch your personal browser’s tabs, history, saved passwords or extensions.

Anthropic’s agent goes by Cowork in its help center, and its pricing page now says “Claude Cowork is now just Claude,” so expect the name to fade as the feature becomes part of the main app. Anthropic says Cowork “uses the same agentic architecture that powers Claude Code, with no terminal required.” It’s available on the paid Pro, Max, Team and Enterprise plans. The full experience is the Claude desktop app for macOS, Windows and Linux (beta), where Claude can work with your local files directly and use a browser built into the app. Sessions run in the cloud, and web and mobile access is rolling out by plan.

You choose how much freedom it gets. Anthropic’s help center describes three modes: Manual, where Claude asks before actions; Auto, where it proceeds and a safety review blocks unsafe actions; and Skip, where it doesn’t pause and nothing checks its actions automatically. Regardless of mode, Claude needs your explicit permission before it permanently deletes files.

If you’d rather have Claude work in your everyday browser, the Claude in Chrome extension can read, click and navigate websites and fill in forms, from a side panel or from Cowork. It’s included on all paid plans. Anthropic’s safety guide lists per-site permissions, a prompt before Claude visits financial sites, confirmations before high-risk actions such as downloading a file or entering sensitive information, and blocked categories such as adult and pirated-content sites. For developers, Claude Pro also includes Claude Code, our pick for agentic coding in the best AI coding assistant.

Not every agent needs to click around a website. Many useful ones read an incoming email, look something up in a spreadsheet, update a CRM and post a summary to Slack. That’s what Zapier Agents is built for. Zapier says its agents can “perform tasks across 9,000+ apps,” use your company knowledge, work on the web and run “on command and while you sleep.”

The free tier is real. According to Zapier’s help center, the free plan includes 400 agent activities a month, with a cap of 10 per run, and Pro includes 1,500 a month with a cap of 40 per run. Zapier’s Agents pricing page lists Pro at about $33.33 a month on annual billing. An activity is one billable step: a trigger firing, a knowledge-source lookup, an action, a web browse or scrape, a web search, or a message through the Chrome extension. Agent activities are billed separately from regular Zapier task usage.

The per-run cap is a useful safety brake. If an agent hits it with work left, Zapier says it asks for your input before continuing, so one confused run can’t burn through your month. Zapier’s AI platform page says it maintains SOC 2 Type II certification and that its subprocessors are prohibited from training their models on customer content.

Google’s agent lives in the browser most people already use. Its help center describes auto browse as a way to have Gemini in Chrome complete multi-step tasks: you describe the task, review Gemini’s plan and click Start Task. Gemini can use your local Chrome, including sites you’re already signed in to, or a remote cloud browser, and it may switch to the remote browser if your device is off. With your permission, Google Password Manager and Sign in with Google can log Gemini in to sites, and Google says it doesn’t share your passwords with Gemini.

Control is straightforward. You can click “Take over task” at any time, and Gemini may hand control back to you for steps such as finalizing a payment, accepting terms of service or creating an account. Google says it asks for confirmation before starting tasks, sending communications, modifying data, submitting forms, scheduling events or accessing sensitive financial or health sites. Google’s subscription page lists Chrome auto browse on Google AI Pro ($19.99 a month) and “highest access” on Ultra, and the help center sets limits of 20 requests a day on Pro and 200 on Ultra.

Lindy is closer to a virtual assistant than to a browser robot. Its pricing page says it manages inboxes and drafts replies in your voice, handles scheduling, meeting prep and follow-up, and works in Slack, iMessage and more than 1,500 integrations. Team plans let you connect several inboxes, depending on the tier.

The approval model fits the job. Lindy says actions with outside impact, such as sending an email, wait for your approval, while read-only lookups from approved sources don’t. On privacy, Lindy’s pricing page says data is encrypted and “never sold, and never used to train models,” and lists SOC 2, GDPR and PIPEDA compliance, plus HIPAA compliance with a signed BAA on Enterprise.