TL;DR. A proactive AI agent works between your chats: it watches the apps you connect, tells you what needs you, takes on jobs you hand it, and asks before it sends, pays, signs or deletes. OpenAI’s dots and Anthropic’s Claude Duet run in their own cloud computers. Start with one job whose output you can read.
A chat assistant waits. Anthropic says of its own Claude app that it “waits for you to open a chat, and you work on something alongside Claude.” Two of the newest products, OpenAI’s dots and Anthropic’s Claude Duet, break that habit on purpose, and Meta’s Muse is pitched as a “personal AI agent.” OpenAI and Anthropic describe what they’ve built in similar terms: an always-on or proactive agent.
A proactive AI agent is an AI that works between your conversations and comes to you. It watches the apps you connect, tells you what needs your attention, carries out tasks you hand it (in the cases of Duet and dots, in its own cloud computer), and stops to ask before it does anything sensitive. In plain terms: an assistant that taps you on the shoulder instead of waiting to be tapped.
Last reviewed: October 22, 2026. Reviewed quarterly, because as of that date Duet is a research preview and dots are rolling out in stages.
What a proactive agent actually is
The phrase comes from the products themselves. Anthropic describes Claude Duet as “a proactive agent that understands what needs your attention,” and Engadget notes that “Like OpenAI, Anthropic is billing Duet as a proactive agent” (Engadget). OpenAI’s own word is “always-on”: its documentation calls a dot “an always-on agent that keeps work moving and asks for your input when a decision needs you” (ChatGPT Learn). Meta calls Muse a “personal AI agent,” according to CNBC (CNBC).
Different words, one idea, built from four parts:
- It comes to you. A normal chatbot starts when you do. Anthropic’s Duet page says the Claude app “waits for you to open a chat, and you work on something alongside Claude,” while Duet “tells you what needs your attention each morning” and “reaches out when one is done or needs your approval” (Claude).
- It works somewhere else. Both Duet and OpenAI’s dots have their own cloud computer and browser, so work continues when your laptop is shut. OpenAI says a dot “can keep working even when your computer is off” (ChatGPT Learn); Anthropic says Duet runs on its own virtual machine with its own browser (Anthropic).
- It remembers you. It uses your history and preferences, so you don’t restart the explanation each time. Our guide to AI memory covers how that works and why it matters.
- It asks before it acts. The point of the design is that the agent does the legwork and you keep the decisions that matter. More on that below.
Proactive agent vs chatbot vs agentic AI
The terms overlap, and people mix them up. Here is how they differ, using what the vendors say about their own products.
| Question | A chatbot | A proactive agent |
|---|---|---|
| Who starts the work? | You, by opening a chat | The agent reaches out, or you message it |
| Where does it run? | In your chat window | In its own cloud computer and browser (Duet, dots) |
| When does it speak up? | When you ask | When it judges something needs you; dots can “decide when to pause and wake up to continue work” |
| What does it do without asking? | Answers | Anthropic says Duet reads, drafts and prepares work on its own and asks before sensitive actions |
Agentic AI is the broader pattern: AI that takes multi-step actions toward a goal. Our guide to agentic AI covers it. A proactive agent is agentic AI plus initiative: it doesn’t only act when told, it notices. And a tool like Claude Cowork is the other model: you describe an outcome and it carries out the work. Whether a given product is proactive or reactive comes down to who starts the conversation.
The loop: watch, decide, ask, act
The vendors’ descriptions line up on one loop, and it is worth knowing because it tells you where to look when something goes wrong.
Watch. OpenAI says a dot “works between conversations. It tracks progress, works out what needs to happen next, and follows through as things change.” Anthropic says Duet checks your connected accounts “every few hours,” because “checking nonstop would use up your plan’s limits much faster” (as Engadget quotes it). “Always-on” therefore describes where it lives, not whether it watches every second.
Decide. This is the new and risky part: the agent has to judge what is worth your attention. Anthropic is plain about the limits, saying Duet “may misread a thread or miscalibrate what’s urgent, and it can miss something that arrives between checking in with you” (Claude). Every’s reviewer, who used it, recommends it to people who do knowledge work, and he also reports that OpenAI’s Dot has “become less proactive” and has lost tasks, a reminder that the products are early.
Ask. Before Duet sends a message, makes a payment, signs a document or deletes anything, Anthropic says it shows you the exact action and waits for approval. OpenAI’s documentation says a dot’s research of connected information is read-only and that “follow-up actions require the applicable permissions.”
Act. After you approve, it carries the action out. OpenAI’s launch post gives an example from an early tester: his dot noticed he had forgotten to invoice a publication, prepared the invoice and sent it after he approved it.
Who has one today
Three products use the idea. They launched between September and October 21, 2026.
| Product | Maker | Launched | Who can get it | How it is described |
|---|---|---|---|---|
| Muse | Meta | September 2026, per CNBC | Not covered here | A “personal AI agent,” marketed primarily to consumers |
| Dots | OpenAI | September 29, 2026 | Rolling out in eligible markets on the Pro, Business Premium and Enterprise plans | An “always-on agent” with its own cloud computer |
| Claude Duet | Anthropic | October 21, 2026 (research preview) | US Pro and Max plans, by waitlist | A “proactive agent” for professionals and small business owners |
CNBC reports that Muse “rocketed to the top of Apple’s App Store” after launch, and that Meta’s CEO told analysts more than 10 million people in the US and Canada “have set up Muse personal agents” over the past month and a half (as of Meta’s October 21 earnings call). For the two business-leaning products, our deeper guides are what Claude Duet is and how to join the waitlist and OpenAI Dots explained. For Meta’s, see what Muse is and whether to trust it.
What it costs beyond the subscription
Background work is not free in terms of what your plan lets you do. Anthropic says Duet “shares your plan’s usage limits with the rest of Claude,” that “its background checks count toward those limits,” and that larger tasks, such as building a financial model or a deck, “require more model turns and background work.” Its help article says usage counts against the plan’s “five-hour session and weekly limits.” Engadget’s reaction was that sharing usage with Claude’s other tools “sounds like a recipe for the new agent to blow through your rate limits.”
I could not find a published figure for what one check or one morning brief uses, from Anthropic or anyone else, so the only way to know is to measure your own week before and after. Our guide to AI usage limits explains how these caps work.
How much control you keep
The list below is mostly about Duet, with OpenAI’s differences noted where I read them. For Duet, the shape is reading and drafting without asking, and approval before it sends, pays, signs or deletes.
- Asks first. Duet shows the exact action and waits before it sends a message, makes a payment, signs a document or deletes anything.
- Never, even with approval. Anthropic lists trading securities, moving money between accounts and accessing someone else’s account.
- Can you look over its shoulder? OpenAI says that when you want to inspect a dot’s browser or do a step yourself, you can open its computer and select Take over. Anthropic’s help article says of Duet: “You can’t watch Claude Duet work in its browser or take over the browser.”
- Credentials. Anthropic says Duet never directly sees your passwords or sign-in tokens.
- Imperfect defenses. Anthropic’s help article says “no defense is perfect, which is one reason consequential actions need your approval.” Anthropic says Duet’s defenses against hidden instructions in what it reads (prompt injection) draw on its browser-use research; our guide to prompt injection explains the risk.
What this means for small business owners
Anthropic and CNBC both say Duet is aimed at small businesses and professionals, and the examples Anthropic gives are office admin: finding unpaid invoices and drafting the reminders, preparing payroll for you to review, and getting contracts ready to send. First action: pick one task you do every week that ends in a draft you can read, such as late-invoice reminders, and make that the agent’s only job for the first two weeks.
What this means for accountants and bookkeepers
An agent that reads your inbox can see client data, and that is a professional-duty question before it is a productivity one. First action: check your confidentiality obligations and your firm’s policy before connecting anything, and keep the agent off client folders until you have decided what it may see.
What this means for freelancers
A freelancer’s work is a long list of small follow-ups, which is the shape these agents are built for. OpenAI’s own example, a dot that noticed a forgotten invoice and prepared it for approval, is a freelancer’s chore. First action: connect only your invoicing app and ask for a weekly list of anything unpaid, with drafts and no sends.
Common misconceptions
“Proactive means it acts on its own.”
Not for the sensitive things. Anthropic describes approval before sending, paying, signing or deleting, and OpenAI’s own example has the dot sending an invoice only after the tester approved it. What the agent does on its own is read, draft and prepare.
“Always-on means it watches everything, all the time.”
Anthropic says Duet checks your connected accounts “every few hours.” A dot “can decide when to pause and wake up.” They are persistent, not continuous.
“It replaces an assistant.”
Anthropic says Duet “may misread a thread or miscalibrate what’s urgent,” and that personal errands such as booking tickets are not “quite as strong,” per CNBC. Treat it as a first-pass sorter that you check.
“It’s just a scheduled task.”
A scheduled task runs when you told it to. OpenAI says a dot doesn’t need “a fixed schedule for every follow-up,” and Anthropic describes Duet as telling you what needs your attention without being asked each time. The initiative is the difference.
See also
Courses, blog posts and related terms about proactive agents.
Courses on proactive agents and adjacent topics (first 2 lessons free)
- OpenAI dots: Delegate to an Always-On Agent: connections, custom rules, schedules and a stop plan
- Claude Cowork Essentials: delegate file work on desktop, web and mobile
- AI Agents Deep Dive: how agents plan, use tools and fail
- AI Fundamentals: how AI works, in plain language
Related terms
- Agentic AI: the broader pattern
- Ambient AI: AI that works in the background of a task
- AI memory: how an agent remembers you
- AI usage limits: why background work counts
- Prompt injection: the main risk when an agent reads your content
- Claude Cowork: the job-description model of an agent
The bottom line
The proactive agent is the idea that you shouldn’t have to start every conversation. Three companies shipped a version of it between September and October 21, 2026. Anthropic and OpenAI each describe the agent doing the legwork and asking you before the actions that matter. That approval step is the part you control, so keep it on.
If you can get access, start with one connected app and one job whose output you can read in two minutes. Keep approvals on. Measure what it costs you in usage limits. And if you want to practise the habits before any of these reach you, OpenAI dots: Delegate to an Always-On Agent (first 2 lessons free) teaches connections, rules, schedules and a stop plan.
Frequently asked questions
What is a proactive AI agent? A proactive AI agent is an AI that works between your conversations and comes to you: it keeps watching the apps you connect, tells you what needs your attention, takes on tasks you hand it, and stops to ask before it does anything sensitive. Anthropic describes Claude Duet as “a proactive agent that understands what needs your attention,” and OpenAI describes a dot as “an always-on agent that keeps work moving and asks for your input when a decision needs you.”
How is a proactive agent different from a chatbot? A chatbot waits for you to open a chat. Anthropic’s own wording for the Claude app is that it “waits for you to open a chat, and you work on something alongside Claude,” while Claude Duet “comes to you.” A proactive agent also runs somewhere you aren’t: Duet and OpenAI’s dots each have their own cloud computer and browser, so they can keep working when your laptop is off.
Which proactive agents exist right now? Three launched between September and October 21, 2026. Meta’s Muse launched in September 2026 and Meta markets it as a “personal AI agent.” OpenAI introduced dots on September 29, 2026. Anthropic opened a research preview of Claude Duet on October 21, 2026. Dots are rolling out gradually and Duet is a waitlisted preview; I did not read Meta’s availability terms, so check availability before planning around any of them.
Does a proactive agent act without asking me? Not for sensitive actions, according to Anthropic’s and OpenAI’s documentation. Anthropic says Duet shows you the exact action and waits for approval before it sends a message, makes a payment, signs a document or deletes anything, and it will not trade securities, move money between accounts or access someone else’s account even with approval. OpenAI says a dot reaches out with decisions that need you, and that follow-up actions require the applicable permissions.
Does a proactive agent use my usage limits? Claude Duet does. Anthropic says it shares your plan’s usage limits with the rest of Claude, its background checks count toward those limits, and larger tasks use more. I did not find a published figure for how much a check uses, from Anthropic or anyone else.
Is a proactive agent safe to connect to my email? It depends on what you connect and what it is allowed to do. Anthropic says that no defense is perfect, which is one reason consequential actions need your approval, and it tells you to review drafts and approval requests before you confirm them. Connect one app for one job first, keep approvals on, and read each action before you approve it.
Sources
- Claude, “Claude Duet product page,” accessed 2026-10-22. https://claude.com/product/claude-duet
- Anthropic, “Introducing Claude Duet,” accessed 2026-10-22. https://claude.com/resources/articles/claude-duet
- Anthropic Help Center, “Get started with Claude Duet,” accessed 2026-10-22. https://support.claude.com/en/articles/17631495-get-started-with-claude-duet
- OpenAI, “Introducing dots,” accessed 2026-10-22. https://openai.com/index/introducing-dots/
- ChatGPT Learn, “Meet dots,” accessed 2026-10-22. https://learn.chatgpt.com/docs/dots
- Engadget, “Claude Duet is Anthropic’s answer to Meta Muse and OpenAI Dots,” accessed 2026-10-22. https://www.engadget.com/2293646/claude-duet-is-anthropics-answer-to-meta-muse-and-openai-dots/
- CNBC, “Anthropic announces preview of Claude Duet, a personal AI agent to take on Muse, Dots,” accessed 2026-10-22. https://www.cnbc.com/2026/10/21/anthropic-preview-of-claude-duet-personal-ai-agent-take-on-muse-dots.html
- Every, “Vibe Check: Duet, Anthropic’s Always-on Agent,” accessed 2026-10-22. https://every.to/vibe-check/vibe-check-duet-anthropic-s-always-on-agent
