AI agents went from demo to product in 2026: nearly every major lab and a wave of startups now sell an assistant that doesn't just answer questions but books, buys, files and follows up on your behalf. The hard part is no longer finding an agent. It's matching one to the job, because they differ sharply in where they run, what they can touch, what they cost, how much they ask before acting, how easy they are to set up and how much upkeep they need. This guide compares nine leading options side by side so you can pick the right tool for the task.
Short answer: choose by where your work already lives and how much you want to control. If you want open source on your own hardware, look at OpenClaw. If you work in Google Workspace, look at Gemini Spark. For Microsoft 365, look at Microsoft Copilot, and if you already pay for ChatGPT Pro or Business Premium, look at ChatGPT dots. For free consumer errands over WhatsApp, look at Muse, and for a text-only life-admin assistant, look at Instinct (invite-only). For file and document work, look at Claude Cowork. For scheduled or event-triggered automations, look at Manus, and for several named agents sharing a cloud computer, look at Grok Bot.
This guide covers agents you message to do tasks for you. Developer frameworks for building your own agents are a separate category and aren't covered here.
Quick answer by use case
| If you mainly want to… | Look at first |
|---|---|
| Run an agent on your own machine, with any model, under an open-source licence | OpenClaw |
| Hand ongoing business work to several named agents that share one cloud computer, browser and logins | Grok Bot |
| Keep an always-on agent inside ChatGPT, reachable in Slack and Teams | ChatGPT dots |
| Try a free consumer agent for shopping, bookings and reminders, including in WhatsApp | Muse (Meta) |
| Text or call an assistant for everyday errands, with no app | Instinct |
| Do research, documents, spreadsheets and decks from your own files | Claude Cowork |
| Work across Gmail, Docs, Slides and other Google apps | Gemini Spark |
| Work across Word, Excel, PowerPoint and Outlook | Microsoft Copilot |
| Set up schedule- or event-triggered workflows that produce finished deliverables | Manus |
Comparison table
| Platform (maker) | Who it's for | What it does | Where it runs | Pricing as listed (Oct 2026) | Safety and approval controls | Integrations |
|---|---|---|---|---|---|---|
| OpenClaw (OpenClaw Foundation, a US 501(c)(3)) | Developers, power users and teams who want to self-host | A self-hosted gateway that connects chat apps to an AI agent with tools, sessions, memory and multi-agent routing | Your own Mac, Windows or Linux machine, or a server | Free, MIT-licensed, with no paid tier. You supply the model and hardware. | Configurable exec approvals (deny, allowlist, ask, auto, full) and optional sandboxing, which is off by default. Default policy on a gateway host runs commands without prompts, so set your own policy. | WhatsApp, Telegram, Discord, Slack, Signal, iMessage and more (29 channels listed). Hosted, subscription or local models. |
| Grok Bot (x.ai) | Individuals and teams handing off multi-step work | Named "Bots", each with memory and a job. They work in parallel, message each other and can learn a workflow from a demonstration. | A persistent cloud computer with a browser, files and a terminal. Apps on macOS, Windows, Linux, iOS and Android. | Included on The X Subscription Premium ($30/mo), Super ($100/mo) and Ultra ($200/mo), but not Plus or Free. Also included with paid Cursor plans. | Returns to you when approval is needed, with drafts you approve before sending. Optional Auto Review checks sensitive actions. Training opt-out. Enterprise admins can set DLP and network controls. | Connectors where available, otherwise computer use in your apps and websites. Slack and GitHub event triggers. |
| ChatGPT dots (OpenAI) | ChatGPT Pro and Business Premium users | An always-on agent that tracks a responsibility, follows up between conversations and can delegate tasks to ChatGPT Work or Codex | Its own cloud computer and browser. Reachable from ChatGPT (mobile, desktop and desktop web), Slack and Teams. You can connect one personal computer. | First dot included with Pro 100, 200 or 500 ($100, $200 or $500 a month) and Business Premium. Pro plans aren't available in the EEA, UK or Switzerland. | Built-in safeguards and app permissions, plus automatic approval checks before actions that affect accounts or share information. Optional custom rules, Take over, and Pause. Stopping doesn't undo completed actions. | Supported ChatGPT plugins (for example Gmail, Google Drive and GitHub) |
| Muse (Meta) | Consumers and small businesses | A personal agent for errands, bookings, shopping, monitoring and goals. It keeps working in the background. | Muse Secure VM, an isolated cloud Linux computer with a browser. Muse app on mobile and Mac, the web, and WhatsApp. | Free with a usage limit. Power is $20/mo and Maximum is $100/mo. Still in limited testing and not available everywhere. | Asks before sending messages, buying or sharing with a connected app, with allow once, always allow or deny. Audit trail of actions. A credential store the agent can't read, and one-time card numbers at checkout. | Email, calendar, Instagram and other connectors. Can build its own tools. |
| Instinct (Spear Street Technology, Inc.) | People who want a text-message assistant for life admin | Handles bookings, orders, appointments and follow-up emails. Uses "a phone and computer in the same way that you do." | You text or call it over iMessage, WhatsApp or email. There's no app. | No published pricing. Invite or waitlist only, and free for invite-only users according to its founder (The Verge, Oct 9, 2026). | Website logins go into a "secure vault" (The Verge). Read its terms and privacy policy for data use and transactions on your behalf before you connect accounts. | Google Workspace, Slack and Notion (The Verge) |
| Claude Cowork (Anthropic) | Knowledge workers and teams | Multi-step research, analysis, documents, spreadsheets and decks in the folders and tools you choose. Can run on a schedule. | Desktop app for local files and apps, with web and mobile in beta. Cloud tasks keep going after you close your laptop. | Included in Pro ($20/mo monthly, or $17/mo billed annually), Max (from $100/mo), Team and Enterprise | You choose which folders and tools it can reach, and deleting needs your approval. Optional permission setting to show its plan and wait. Asks before using each app with computer use. Admin controls on Team and Enterprise. | Connectors and plugins (skills, connectors, sub-agents). Can run through Amazon Bedrock, Google Cloud or Microsoft Foundry. |
| Gemini Spark (Google) | Google Workspace and Gmail users | A 24/7 agent for recurring tasks, triggers, skills and workflows across Google apps | Runs in the cloud. Gemini app, plus the Gemini app for Mac. | Requires Google AI Pro ($19.99/mo in the US) or Ultra ($99.99 or $199.99/mo). Country availability varies. | You choose whether to turn it on and which apps it connects to. Designed to ask first before high-stakes actions like spending money or sending email. | Gmail, Docs, Slides, Sheets, Keep and Tasks. Connected apps such as Canva, Instacart, OpenTable, Dropbox and Zillow. Custom MCP servers. |
| Microsoft Copilot (Microsoft) | Microsoft 365 users | Copilot Tasks plans and carries out goals, now or on a schedule, including browser tasks. Microsoft is moving personal accounts to Copilot Cowork. | Copilot app and web, plus the Word, Excel, PowerPoint and Outlook apps | Copilot is in Microsoft 365 Personal ($9.99/mo) and Family ($12.99/mo). Microsoft 365 Premium was $19.99/mo in the US at launch. | Asks for approval or hands control back for purchases, sending personal information, messages, account changes and sensitive submissions. Pause, stop, take control. | Microsoft 365 apps and connectors you authorize (for example email and cloud storage) |
| Manus (Manus AI) | Individuals and teams automating research and deliverables | Agent tasks and Automations started by a schedule, an app event or a plain-language routine | Web app (cloud) | Free (300 daily credits). Pro from $20/mo (4,000 credits). Team from $20 per seat a month. | Test run before going live. "Approval required" setting for anything that sends or publishes. Run history in Updates. | Connected apps such as HubSpot, Gmail, Slack, Notion, GitHub and Shopify, plus webhooks |
"Pricing as listed" means the subscription price on the maker's page. Most plans also have usage limits or credits, and heavy agent work can use them up faster than chat does.
Ease of use and maintenance required
Two practical questions decide whether an agent sticks: how quickly you can get to a first useful task, and how much ongoing care it needs once it's running. Self-hosted tools give you the most control and need the most upkeep. Hosted agents hand most maintenance to the vendor but still need you to keep connections current and answer approvals.
| Platform | Ease of use (getting started) | Maintenance required |
|---|---|---|
| OpenClaw | The most setup. You install the gateway on your own machine or server, connect a model and your chat apps, and set your own approval and sandbox policy. Best if you're comfortable in a terminal. | The most upkeep. You run the host, apply updates, manage model keys and usage costs, and keep your security policy current, since approvals and sandboxing aren't strict by default. |
| Grok Bot | Sign in on an eligible X Subscription or paid Cursor plan and describe the job in chat. A Bot can also learn a routine from a demonstration. | The vendor runs the cloud computer. You keep the app logins and connectors on it current, and you review the drafts and approvals it brings back. |
| ChatGPT dots | Nothing to install if you're already on ChatGPT Pro or Business Premium. You give a dot a responsibility in ChatGPT. It's still rolling out gradually. | Vendor-hosted. You keep plugin connections and any custom rules up to date, and you answer the decisions it brings you. |
| Muse (Meta) | Chat in the Muse app, on the web or in WhatsApp, with a free tier. It's in limited testing and not available everywhere. | Vendor-hosted. You answer allow-once, always-allow or deny prompts and watch the weekly usage allowance. |
| Instinct | You text or call it, with no app to learn. Access is by invite or waitlist. | Vendor-hosted, with little public documentation. You keep the logins in its vault current. With no published pricing, there are no published plan terms to track yet. |
| Claude Cowork | A desktop app where you pick the folders and tools it may use. It's included in paid Claude plans, and web and mobile are in beta. | You review scheduled tasks and the folders and connectors you've granted, and you keep the desktop app updated for local-file work. |
| Gemini Spark | You turn it on in the Gemini app if your plan and country qualify. It's easiest if your work already lives in Google apps. | Vendor-hosted. You review connected apps and triggers. Any custom MCP servers you add are yours to maintain. |
| Microsoft Copilot | Built into the Copilot app and the Microsoft 365 apps. You give it a goal to run now or on a schedule. | Vendor-hosted. Microsoft says it's moving personal accounts from Copilot Tasks to Copilot Cowork, so expect a migration. |
| Manus | A web app. You describe a routine in plain language and do a test run before it goes live. | You watch run history and credits (300 a day on Free), and you fix connected apps and webhooks when they break. |
This is our reading of the setup and upkeep each maker documents, using the sources linked in this guide. We didn't run a hands-on usability test, and your experience will depend on your accounts and how much you connect.
Short profiles
OpenClaw
OpenClaw is open-source software you install yourself. One gateway connects your chat apps to an agent that can browse, fill forms, read and write files and run commands. It has no hosted tier and no paid version, and it's stewarded by the independent OpenClaw Foundation. It suits people who want their data and model choice on their own machines and are comfortable setting security policy. Sandboxing and command approvals are configurable, but they aren't strict by default. Sources: openclaw.ai, docs, exec approvals, sandboxing.
Grok Bot
Grok Bot gives each Bot a persistent cloud computer with a browser, files and a terminal, so work continues while your laptop is closed. All your Bots share that computer, including its files and app logins, so treat anything you put on it as available to every Bot. Bots can run in parallel, pass work to each other and learn a routine from a demonstration. It's sold through The X Subscription (Premium and up) and is included with paid Cursor plans. Sources: x.ai/bot, docs, pricing.
ChatGPT dots
A dot is OpenAI's always-on agent inside ChatGPT. You give it a responsibility, and it keeps track, decides when to follow up, and brings you decisions. It has its own cloud computer and browser, and you can reach it in ChatGPT, Slack, Teams or by voice call. Dots are rolling out gradually, and the consumer Pro plans exclude the EEA, UK and Switzerland. Sources: Meet dots, getting started, Pro tiers.
Muse
Muse is Meta's personal agent for everyday tasks such as reservations, shopping, travel, reminders and price monitoring. It runs in an isolated cloud computer you can also use yourself, and you can chat with it in its app or in WhatsApp. It's free with a usage limit, and paid plans raise the weekly allowance. Meta says it's in limited testing and not available in every location. Sources: ai.meta.com/muse, Muse subscriptions.
Instinct
Instinct is a personal assistant you text or call, with no app. Its maker says it can use a phone and computer the way you do, for jobs like ordering groceries or booking a train. As of October 2026 it was invite-only or by waitlist, with no published price. Because it can act on your accounts, read its terms and privacy policy first. TechCrunch summarized their scope in August 2026. Sources: instinct.com, The Verge.
Claude Cowork
Cowork is where you hand Claude multi-step knowledge work. It works in the folders and tools you choose, uses a built-in browser for web steps, and can run scheduled tasks unattended. It's included in paid Claude plans. Use the desktop app to work with local files. Sources: claude.com/product/cowork, pricing.
Gemini Spark
Spark is Google's 24/7 agent in the Gemini app. It's closely tied to Workspace and supports recurring tasks, triggers, skills and custom MCP connections. It requires a Google AI Pro or Ultra plan, and availability depends on your country and language. Sources: Google blog, what's new, Google AI plans.
Microsoft Copilot
For personal Microsoft accounts, Copilot Tasks lets you hand Copilot a goal to run now or on a schedule. Microsoft documents which actions always need your approval and warns about prompt injection. Microsoft says it's moving these users to Copilot Cowork. Sources: Using Copilot Tasks, plans, Microsoft 365 Premium launch.
Manus
Manus runs agent tasks and Automations that start on a schedule, on an event in a connected app, or from a routine you describe. Its guidance is to use a test run and keep "Approval required" on for anything that sends or publishes. Sources: Introducing Automations, pricing.
How to choose: a checklist
- Availability: confirm your country, age and plan qualify. Several agents are region-limited or rolling out gradually.
- Where your data lives: on your own machine, or on the vendor's cloud computer. Who else, including your other agents, can see what's on it?
- Accounts you'd connect: list them, and connect only what the first job needs.
- Approval defaults: check what it does without asking. Turn on approvals for sending, buying, deleting and sharing before the first real task.
- Audit trail and stop: can you see every action and pause it? Remember that stopping usually doesn't undo what's already done.
- Prompt injection: any agent that reads email or web pages can be fed hidden instructions. Test that first with the prompt-injection checklist.
- Ease of use: how long does it take to get from sign-up to a first useful task? Could someone non-technical on your team set it up, and correct it when it gets something wrong?
- Maintenance required: who handles hosting, updates, expiring logins and broken connectors? Budget a weekly look at run history and pending approvals.
- Real cost: subscription plus usage limits or credits, plus your review time. See how to estimate cost per run.
- Terms and training: read the data-use and model-training settings, and the defaults for personal versus business plans.
- Exit plan: how do you revoke access and delete stored data?
- Pilot small: run one low-stakes, reversible task for a week before you hand over anything important.
If you're rolling one out at work
Agent Field Kit has templates for the checks above. The MCP & Tool-Access Review Kit covers deciding what an agent may touch. The Prompt-Injection Test Pack covers testing it, and the AI Vendor Security & Data-Handling Questionnaire covers vetting the vendor. The Cost & ROI Calculator helps you budget it, and the Agent Security Audit Kit helps you audit the whole setup. Templates start at $29, with a 14-day refund.
Which agent do you use? Take our one-question poll and see live results after you vote.
Sources
All accessed Oct 18, 2026. OpenClaw: openclaw.ai, docs.openclaw.ai, exec approvals. Grok Bot: x.ai/bot, overview, pricing. ChatGPT dots: Meet dots, Pro tiers, business pricing. Muse: ai.meta.com/muse, subscriptions. Instinct: instinct.com, The Verge, Oct 9, 2026, TechCrunch, Aug 24, 2026. Claude Cowork: product page, pricing. Gemini Spark: Google blog, May 19, 2026, what's new, plans. Microsoft Copilot: Using Copilot Tasks, plans. Manus: Automations, pricing.
FAQ
Which AI agent platform is best?
There is no single best one. Pick by where your work already lives (Google, Microsoft, ChatGPT, chat apps or your own hardware), which regions and plans it is available on, and how its approval controls handle sending, buying and deleting.
Is there a free AI agent?
OpenClaw is free, open-source software you run yourself (you supply the model and hardware). Muse and Manus have free tiers with usage limits, and Instinct was free for invite-only users as of Oct 2026, according to its founder.
Are these prices current?
They are the listed prices we found on each maker's own pages on Oct 18, 2026, in US dollars. Plans change often, so check the linked pricing page before you buy.