What are personal AI agents? Muse, dots, Gemini Spark and Grok Bot explained

In the last five months four of the largest AI companies have shipped the same kind of product. Google announced Gemini Spark in May. xAI launched Grok Bot in August. Meta launched Muse in September, and OpenAI followed three weeks later with dots. Each one is described as a personal agent: software that remembers you, works on its own computer while you're away, and acts for you in your accounts. They aren't only coming from the big labs. Instinct, a startup you reach by text message, raised $1 billion in September at a $10 billion valuation.

That changes who turns up at a company. A personal agent doesn't only answer questions about an airline, a bank or a shop. It opens their website, signs in, fills in the forms and asks you to approve the result. This post explains what a personal agent is, how it differs from the chatbots and agent tools people already use, and how the companies those agents visit can work with them.

#What is a personal AI agent?

A personal AI agent is an AI assistant that keeps working for one person over time: it remembers what that person has told it, connects to their accounts, runs on its own computer in the cloud, and can act and make suggestions without being asked, checking with the person before anything sensitive such as a payment.

Meta's own definition is short: "Muse is a personal AI agent. It doesn't just answer questions, it actually does the work."

Four traits come up in every launch:

• Memory. It learns from what you tell it and carries that into the next task. Meta says Muse can act on details you mentioned only once, and you can tell it to forget things.
• Its own computer. It runs on a cloud machine with its own browser, so it keeps going when your laptop is shut. Muse runs on "Muse Secure VM, a dedicated secure computer with its own browser". OpenAI says, as quoted by Engadget, that each dot has "their own cloud computer, their own browser, and the apps you've connected".
• Proactivity. It doesn't wait to be asked. Muse "can make suggestions unprompted". Google says Spark is "designed to proactively manage tasks".
• Approvals. It stops before anything that matters. Muse "checks with the person before sensitive actions like sending an email or making a purchase". Spark asks "before performing high-stakes actions like spending money". Grok Bots "only come back when something needs your approval".

#Chatbots, agent tools and personal agents

It helps to put personal agents next to the tools people already use.

A chatbot handles one request at a time. Upload a report and it explains that report. It doesn't connect it to last year's.

An agent tool you set up yourself keeps notes on you, usually as plain text files, and reads them when a task needs them. It can work across your folders and connected apps. The trade-off is that you design the system: what it stores, where, and what it runs. Open-source projects such as OpenClaw and Nous Research's Hermes Agent take this route, and so do coding agents used for everyday work.

A personal agent is the packaged version, with one addition: it is always on. You give it a goal ("sleep better", "find a flat", "chase overdue invoices") and it builds the plan, keeps the notes, checks back during the day and comes to you when it needs a decision.

The trade-offs follow from that. A packaged agent is easier to start but organised the vendor's way. The more it knows about you, the harder it is to move to another one. And it holds a lot: Muse asks for access to email, calendars, payments, and health and fitness apps, which is why TechCrunch headlined its launch coverage "Will consumers trust it?"

#The main launches

They differ in emphasis:

• Muse is aimed at everyday life. Meta lists sending emails, booking travel, lowering bills and filling out forms. A separate Sentinel agent runs on the same machine, kept apart from Muse, and Muse "can't reach the internet unless Sentinel approves".
• dots sit inside ChatGPT and connect to more than 4,000 apps through OpenAI's plugins. Background research uses read-only tools that, per Engadget, "can't send messages, change app content, or control your browser or computer". Dots "start with built-in rules for when to act independently and when to ask for approval", and users can add their own.
• Gemini Spark is tied to Gmail, Docs, Calendar and Drive, uses Chrome for web tasks, and names Canva, OpenTable and Instacart as early partner apps.
• Grok Bot is pitched as a team of "AI teammates" that sign into your tools, remember conversations and work in parallel.
• Instinct has no new interface: you text it, and it "can use a phone and computer in the same way that humans do". Its examples are disputing a bill, sending birthday gifts, ordering groceries and booking a doctor's appointment. TechCrunch reports a concierge feature that phones businesses with no online booking, and a "trusted person network" that lets one user's agent make plans with a friend's agent.

#How a personal agent reaches a company

Most of what a personal agent does for you happens at someone else's company: your airline, your bank, your energy supplier, a shop. Today there are two ways in.

1. A browser, as you. The agent opens a real browser on its cloud computer, signs in with your details and uses the site the way you would. This works on any website and needs nothing from the company. It is also where the friction is. On 20 September Amazon began blocking Muse, telling users that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use". Amazon said it hadn't been told Muse would access its store, that the agent doesn't identify itself when it browses, and that it appears to capture customer credentials. Meta says Muse has no visibility into passwords or payment methods. Search Engine Journal pointed out that Meta's crawler documentation lists five user agents and Muse isn't one of them, so a robots.txt rule had nothing to name. A day later Shopify announced checkout for Muse across its merchants. Two companies, the same week, opposite choices, and in both cases the question underneath was the same: who is this agent, and what is it allowed to do?

2. A published interface. The company tells agents how to work with it, and the agent uses that instead of pages built for people. For buying, commerce protocols such as the Universal Commerce Protocol define catalog, cart and checkout. For the wider relationship (who the agent is acting for, signing in to the person's account, what it may see and change) Meta and Sierra published the Personal Agent Protocol on 9 October.

#Where the Personal Agent Protocol fits

The Personal Agent Protocol (PAP, nicknamed "Poppy") describes today's choice as "all or nothing": either the agent passes as the user, "signing in with their password and clicking through pages built for people", or the company blocks it. Draft 0.1 sets out the range in between.

• Discovery. A company publishes /.well-known/poppy.json saying how personal agents can work with it: its sign-in server, the scopes it offers, and which routes it supports (its own agent, its APIs, or its website).
• Identity. The agent identifies itself with a client_id that is a URL to its own metadata and public keys, so the company knows which agent it is dealing with.
• Permission. The person signs in on the company's own page (or on another device with a code) and grants scoped access, such as poppy:read to see orders or bookings and poppy:write to change them. Session tokens are short-lived and bound to a key the agent holds.
• One session. API calls, browsing and conversation with the company's own agent all count as one visit.

The spec's own example is an airline. Any agent can search flights. A signed-in agent can see the user's trips. Buying a ticket needs the user's approval. The user, for their part, might let an agent see trips but not change them.

Draft 0.1 leaves payments, push notifications and attachments as open topics, and says any part can change before a stable version. The partner list names three personal agents: Meta, which co-authored the protocol, OpenAI and Instinct. Google and xAI aren't named.

#What it means for companies

• Agents are already customers. A personal agent working for someone is that person, with their intent, often an existing account holder. Treating every automated session as a bot to block will turn some of them away.
• Decide your policy on purpose. Amazon and Shopify made opposite choices in the same week. Both are policies. What matters is that yours is written down and that agents can read it, rather than learned from a popup.
• Say how agents should work with you. A poppy.json is a small JSON file that points agents at your sign-in server, your scopes and your APIs. If you already run OAuth for customer accounts and an API or MCP server, most of it exists already. The guide walks through every field, and the validator checks a draft before you publish it.
• See who has started. The sites list shows the domains publishing a poppy.json, and the checker reads any domain's file against draft 0.1.

#What we'll watch

• Whether Meta, OpenAI, Google and xAI document how their agents identify themselves to websites.
• Which personal agents implement the Personal Agent Protocol as clients, starting with the three it lists: Muse, dots and Instinct.
• How Amazon and Meta resolve the Muse dispute, and whether other large companies follow Amazon or Shopify.
• The Personal Agent Protocol's path from draft 0.1 to a stable version, including how it handles payments.
• Availability outside the US. Most of these launches started there.

#FAQ

#What is the difference between a chatbot and a personal agent?

A chatbot answers what you ask, one conversation at a time. A personal agent remembers you between tasks, connects to your accounts, runs on its own computer in the cloud, works when you're not there, and suggests things without being asked. It still asks for your approval before sensitive actions such as a purchase.

#What are examples of personal AI agents?

The main launches of 2026 are Meta's Muse (September), OpenAI's dots (September), Google's Gemini Spark (announced in May), xAI's Grok Bot (August, beta) and Instinct (August, invite-only). Open-source projects such as OpenClaw and Hermes Agent let people run a similar agent themselves.

#Can a personal AI agent buy things or change my bookings?

Yes, with your approval. Muse checks out with Link by Stripe and asks before making a purchase. Gemini Spark asks before spending money. Whether the agent can finish the task depends on the company: some allow agents, some block them, and some publish a way for agents to work with them.

#Why did Amazon block Meta's Muse?

Amazon said it wasn't told Muse would access its store, that the agent doesn't identify itself when browsing, and that it appears to capture customer credentials, which Amazon said breaks its Conditions of Use. Meta says Muse has no visibility into passwords or payment methods.

#What is the Personal Agent Protocol?

An open protocol, launched by Meta and Sierra with draft 0.1 published on 9 October 2026, for how a person's AI agent works with a company on their behalf. A company publishes /.well-known/poppy.json; the agent identifies itself, the person signs in and grants limited access, and the agent works through the company's APIs, website or own agent within that access.

#Do personal agents work when my computer is off?

Yes. Muse, dots, Spark, Grok Bot and Instinct each run on a computer the vendor provides, so they keep working when your laptop is closed or your phone is locked.

#Are personal AI agents safe to give access to my accounts?

Each vendor describes its own safeguards: Meta's separate Sentinel agent and isolated VM, OpenAI's read-only background research and approval rules, and Google and xAI's approval steps. These are vendors' descriptions. You choose which apps each agent can reach and can disconnect them. Giving any agent access to email, payments or health data is a trust decision worth making app by app.

#Sources

• Meta, Introducing Muse, a personal AI agent, Meta Newsroom, 8 September 2026 (updated 30 September)
• TechCrunch, Meta debuts its Muse AI agent. Will consumers trust it?, 8 September 2026
• Engadget, Dots are OpenAI's new personal agents, and soon you'll be able to control several of them, 29 September 2026
• TechCrunch, OpenAI launches Dots, its bubbly agentic avatar, 29 September 2026
• Google, The Gemini app becomes more agentic, delivering proactive, 24/7 help, Google blog, 19 May 2026
• xAI, Introducing Grok Bot, 11 August 2026
• Instinct, instinct.com, accessed 14 October 2026
• TechCrunch, Viral AI agent Instinct raises $1B Series C at a $10B valuation, 28 September 2026
• Search Engine Journal, Amazon blocked Meta's Muse and robots.txt had nothing to say, September 2026
• Personal Agent Protocol, draft 0.1 specification, 9 October 2026, and Path to 1.0, 12 October 2026

The Personal Agent Protocol is a draft at version 0.1 and may change. PAP Checker is an independent checker; it isn't affiliated with the protocol's authors.

Related: What is poppy.json? · Check your poppy.json before you publish it · Sites publishing poppy.json