You’ve probably logged on to ChatGPT, Gemini, Claude or another AI chatbot to ask questions. Maybe you’ve asked for detailed research reports. But for most people, the notion of using an AI “agent” — autonomous AI doing work in the background — still seems complex and distant, only for the most advanced users.
In recent weeks, AI companies have been working to change that. A new wave of AI agents aim to give people ways to outsource work to bots to help manage life — all inside a simple chat or messaging app.
Last month, Meta launched Muse, which has quickly become the most downloaded app in the country. In late September, OpenAI debuted Dots for ChatGPT, followed last week by Anthropic’s announcement of Claude Duet. More are on the market from Google, xAI/Grok, Perplexity and others.
While powered by the same type of large language models as popular chatbots such as ChatGPT and Claude, these agents are designed to be much more proactive. They can help people search, shop, book travel, plan meals and manage family schedules. An agent can work in the background, remember preferences, monitor information and handle tasks across various websites and apps.
Let’s say you remember to book something but don’t have time to do it at the moment. You can have an agent try on your behalf. I recently had Muse schedule an appointment with the DMV and then book our dog a haircut. Using the groomer’s website, it found times, sent me options and picked a slot within my chosen window. Could I have done both myself? Sure. But letting Muse do the legwork lightened the mental load.
In the background, the agents are visiting websites and working on your behalf, using log-in credentials or credit cards you give them. For the most part, success depends on whether a website lets an agent in. Amazon, for instance, has blocked agents from shopping on its site. (Agents also have ways to hand over control for a human to manually take action and make changes.)
“The AI word of the year was proactive, and these agents are the proof,” said Allie K. Miller, an angel investor and CEO of Open Machine, an AI advisory firm for Fortune 500 companies. “The goal is actually to never have to prompt anymore. The best way to think of this new generation of AI assistants is ‘ongoing responsibility.’”
Miller has been putting that concept to the test by handing agents a range of tasks. She recently had OpenAI’s Dots choose a friend and plan an afternoon in New York City with a $500 budget for activities, reservations and transportation. Dots successfully planned several activities they hadn’t thought of, including a free rubber-stamp workshop, visiting Chelsea art galleries and eating lobster tacos at Kahlo. But it also repeatedly emailed one restaurant from Miller’s personal account, emailed friends it knew were sick or out of town, forgot to send reminders and wrongly assumed she was gluten-free.
“I had to send more apologies for the AI’s behavior than it would have taken for me to just plan the afternoon myself,” Miller said. “But a worthy experiment nonetheless.”
While there are some common themes for how people are using agents, they also vary in where they’re accessed, what they cost and how they connect to third-party sources. For example, Muse lives across a new stand-alone app, website and other Meta platforms. Meta also plans to add it to Ray-Ban smart glasses and sell devices, such as a Tamagotchi-style device called Muse Charm and another device for connecting home devices. For the tinkering types, Meta also released open weights for Muse’s code to help people connect it to a range of other devices.
Here’s a quick primer on the leading agents and tips for how to start.
- Muse: One of the most popular in recent weeks has been Meta’s Muse, which is free but has paid tiers for higher usage. It’s also among the more user-friendly, with an evolving “Ideas” tab full of suggestions and a “Goals” tab that tracks task progress. Behind the scenes, Muse uses a collection of digital notes and instructions to guide its personality, actions and memory, all of which can be reviewed, edited and deleted.
- ChatGPT Dots: Debuting just a few weeks after Muse, Dots is designed to be a proactive assistant that tracks connected apps. It’s only available for the ChatGPT Pro tier ($100 per month) or Business Premium, making it less worth trying for now unless you’re already a power user. Dots lives within ChatGPT and doesn’t have its own app, but there’s a way to “call” it to give verbal instructions or talk through a task. The Washington Post has a content partnership with OpenAI, the creator of ChatGPT.
- Instinct: Another agent gaining traction is Instinct, which takes a much more minimalist approach of being accessed through SMS and WhatsApp. That’s handy for people who spend a lot of time in text messages or who want a way to interact without learning yet another app, but it might also feel too minimal for some. You can also ask Instinct to create AI-generated voice messages instead of a wall of text. It’s currently free in beta, but requires an invite or being put on a wait list for access.
- Gemini Spark: Google’s agent debuted this spring and since then has expanded access, added more connections to a range of sources, and gained new visibility as part of the Gemini app’s redesign. Spark can help by tracking online prices, setting up automated search queries and monitoring apartment listings, and reviewing subscriptions.
- Claude Duet: The newest agent is Claude Duet, a stand-alone app and website from Anthropic that debuted last week in research preview. The company says it’ll be a business-focused app with more than 6,000 connections and plug-ins that can remember context and help with a range of business tasks. For now, there’s a wait list for U.S. users at the paid Pro and Max tiers only.
How people are using agents
Several themes are emerging from people testing agents in their personal and professional lives. Families are turning agents into digital “chiefs of staff” that combine calendars, school emails and extracurricular schedules to create daily personalized agendas. Agents can also flag overlooked deadlines, permission slips and scheduling conflicts.
For shopping and travel, an agent can help monitor options and compare prices. It can also alert when it finds a deal or even book and buy on your behalf, if a site allows it. People have used Muse to hunt for niche photography gear, shop for new car tires and guide them through Burgundy’s vineyards. But access varies, with major websites like Amazon and Delta blocking agents even as Shopify and Walmart partner with them.
Work-related tasks include setting up automatic monitoring for information across connected services such as Google, Microsoft, Slack and Notion. Agents can also help people with hobbies and personal goals and interests by connecting agents to health and fitness apps, smart home devices and financial services.
Some tips for starting and staying safe
As agents connect to more accounts, the more useful they can be. But there’s a trade-off:When agents have unfettered access, there’s a risk of exposing personal or sensitive information. Before buying or booking something with an agent, be careful what else it could see. Using Muse for our dog’s groomer only revealed his weight, vaccination status and previous visits. Using an agent to book a human doctor through a medical portal could be much riskier. That’s worth noting, since several examples Meta gives for using Muse include tasks such as booking doctor appointments and checking insurance coverage.
“[Consumer agents] can save enormous amounts of time, but that convenience relies on broad access to a person’s digital life,” said Andrew Frank, a Gartner research analyst focused on emerging tech, marketing and data privacy. “People need to understand what economic role their data plays in the service and how much control they have over what their agents remember and what happens to that information. It’s tempting to ignore the fine print, but overlooking the provider’s privacy and deletion policies could lead to unexpected complications.”
New security and privacy risks from agents are already emerging. Last week, researchers discovered thousands of shareable Muse links that exposed private chat logs and documents, leading Meta to suspend the feature until it could be fixed. Major U.S. banks reportedly are also blocking third-party agents after accounts of them taking action beyond users’ instructions. Meanwhile, companies like 1Password are adding new systems for approving agent log ins without exposing user passwords.
- Start small and start slowly. Ask the agent to “watch” rather than buy or book anything. For example, I connected Muse to Spotify to set up a daily scan for concerts I might like. It built a list of around 70 musicians and 20 small-to-midsize local venues. Muse told me it searches 1.5 seconds apart on Ticketmaster to avoid getting throttled.
- Protect personal data. Before connecting an account, check the agent’s privacy settings and data-retention policies. Muse and Instinct automatically opt users in to allow data AI model training, so you’ll want to turn those off. Start with one connection at a time and know that disconnecting an app doesn’t necessarily erase information already collected. Also make sure an agent asks before it sends, buys, deletes or shares anything.
- Protect your money. Before buying or booking anything, make sure to keep final approval, set spending limits and check a merchant’s cancellation and return policies. Meta’s terms already make clear that users are responsible for anything agents spend. Remember not to enter passwords or credit card numbers into AI chats and instead only use secure platforms. Muse, Instinct and Claude Duet all handle payments through Stripe’s Link service, which can generate single-use virtual cards to keep card details hidden.
The post How to get started using AI agents even if you never imagined it appeared first on Washington Post.
