Personal AI agents, compared

In the last few months every big AI company has shipped an agent that does things for you: it reads your email, books, buys, and keeps working after you close the app. They sound the same in a launch video, but they're pretty different once you look at the price, where you talk to them, and what they're actually good at.

Here's how Meta Muse, ChatGPT, Gemini Spark, Grok Bot and Claude compare as of 27 Sep 2026, plus what OpenAI teased this week. These products change weekly, so treat this as a snapshot and I'll keep it updated.

The short version

Every one of these asks before it sends a message or buys something by default. Prices are US monthly as of 27 Sep 2026.

Each one, in plain English

Meta Muse

Meta's personal agent, launched September 8. It runs on its own computer in Meta's cloud and connects to your email, calendar, shopping and payments. It's the most consumer-focused of the group: returns, bills, grocery lists, booking things. You can message it in WhatsApp, and Meta says it's getting its own email address. It's also the only one with a real free tier, and it's been hugely popular: CNBC reported 2.5 million downloads in its first two weeks. The catch is that it's new. Reviewers have reported mistakes like expired discounts and wrong delivery dates, and Amazon blocked it from shopping on its site. My Muse setup guide covers the safe way to start.

ChatGPT Work

OpenAI's agent, launched in July, is built for getting a piece of work finished. You give it a goal and it works for hours to hand back a document, spreadsheet, deck or website. It runs in the cloud, or on your computer through the ChatGPT desktop app's built-in browser, and it asks before using a new site or app. ChatGPT also has scheduled tasks that can run on a timer or kick off when an email arrives. It's built for work you'd otherwise make yourself. If you want it to build things from a folder of files, look at Codex, which shares the same usage allowance.

Gemini Spark

Google's agent works in the background even when your phone and laptop are off, and it's strongest if your life runs through Gmail, Calendar, Drive and Docs. Since August you can set up long-running jobs just by talking to Gemini Live. Connections to your Google apps are off until you turn them on, and it asks before sending, buying or submitting forms. It needs Google AI Pro ($19.99) or Ultra, and it isn't available in the EU or the UK.

Grok Bot

xAI's agent runs on its own cloud computer, signs into your tools, and handles multi-step work across apps and inboxes, coming back when something needs your approval. You can add "ask first" rules for sending, publishing or deleting. It's still in beta, and since August 26 it's included in every SuperGrok plan. I walk through a real Grok setup in setting up Grok in one day.

Claude

Claude doesn't have a Muse-style agent for personal errands, but since September 16 the part that can open your files, use your connected apps and run longer tasks is built into every conversation. Longer tasks and scheduled tasks keep running in the cloud after you close your laptop, and it asks before it takes actions by default. It's the one I use most for work, and my Claude setup guide shows how.

One more worth knowing: Perplexity's agent, Computer, takes tasks by email. You forward something to it and the result comes back in the same thread.

What OpenAI teased this week

OpenAI hasn't launched its answer to Muse and Grok Bot yet, but it's clearly coming. On September 26, OpenAI posted a teaser, and a line reading "o, your always-on assistant" showed up on the ChatGPT Pro upgrade screen for some users. Code in the app also points to it getting its own email address, like Muse. Always-on means it keeps working after you close the chat.

Everything else is still rumor: which plans get it, what it costs, and a reported $500 a month plan. OpenAI's DevDay keynote is Tuesday, September 29, and it's free to watch live (details on my events page). I'll update this page once it's official. Until then, I wouldn't upgrade a ChatGPT plan just to get it.

How I'd choose

You want help with personal errands and bills. Start with Muse Free, read-only, and add access slowly.

You live in Gmail and Google Docs. Gemini Spark, if you're already paying for Google AI Pro or were thinking about it.

You want finished work, like a report, a deck or a site. ChatGPT Work or Claude. You probably already pay for one of them, so try it there first.

You want something that watches a job every week. Grok Bot or a scheduled task in ChatGPT or Claude. The tool you already pay for is usually good enough.

You're not sure. Give two of them the same small job and compare. Most people don't need more than one agent.

Test two agents on the same job

Paste this into two agents, one at a time, and compare what comes back. It's read-only, so nothing gets sent or bought.

Look through my email from the last 30 days and find anything I still
need to act on: bills due, renewals, replies I owe, and appointments.
List each one with the date, what I need to do, and the email it's from.
Mark anything you're unsure about. Don't reply to, send, buy or change
anything. Just give me the list.

Then score them on the questions that matter:

Agent: [ ]
[ ] It found the items I would have found myself
[ ] Nothing on the list was made up or wrong when I checked
[ ] It told me when it wasn't sure
[ ] Setup took less than 30 minutes
[ ] It works with the apps I actually use
Score: __ / 5