The year of the agent showed up late

A version of this article originally appeared in Quartz’s members-only Weekend Brief newsletter. Quartz members get access to exclusive newsletters and more. Sign up here.

In January 2025, Sam Altman wrote a blog post predicting that "in 2025, we may see the first AI agents 'join the workforce' and materially change the output of companies." Plenty of agent tools shipped that year. Outside of coding, none broke out, because they were too complicated for most people to set up.

This month made a better case for 2026 being the year of the agent. The breakout tool was easy enough for nearly anyone with an internet connection to use. 

It just wasn't made by OpenAI.

Muse shopped, booked travel and filled out forms, with a free tier and paid plans up to $100 a month. Walmart $WMT, Sephora, Expedia $EXPE and Shopify $SHOP signed on as partners, and Mark Zuckerberg said Meta would take a commission on purchases.

Problems, however, followed quickly. Amazon $AMZN blocked Muse from its store and said it got no notice and never authorized the agent. Security researchers questioned how much personal data Muse needed, and whether people who aren't tech-savvy understood the risk or how the technology worked.

The worst case (so far) involved a consumer tech reviewer named Matt Robb, who let Muse manage his Facebook Marketplace listings. When a buyer asked about a keyboard, Muse replied as Robb, agreed on a price and sent his Toronto home address. The buyer showed up with his wife and daughter while Muse, still writing as Robb, texted "I'm right here, like waiting for you." Robb was not there.

He had clicked "Allow Always" on a permissions prompt, expecting offers to still reach him for approval. They did not. Meta said it would make the prompt clearer, though David Singleton, who runs its Superintelligence Labs, said reviews of similar reports found Muse had followed direct instructions.

Tuesday was OpenAI's turn. At DevDay it introduced Dots, always-on agents with a cute mascot (not too much like Muse's), their own cloud computers and connections to more than 4,000 apps. Users could message them in ChatGPT, Slack $WORK, and Teams, with texting promised later. From stage, Altman said Dots gave him some of his attention back. Even on his big day, Altman got questions about Muse. He called it a nice product and said he was not threatened by Zuckerberg.

At DevDay, Altman said he expects AI liability to become a “multi-level” system, similar to the auto industry, with responsibility depending on whether the fault is in the model, how it was used, or whether someone intentionally misused it.

There is also the cost to the people running the agents. Coding agents use up to a thousand times more tokens than chatbots, according to one study, and there is little reason to expect general-purpose agents like Muse to be much leaner. That bill lands on users or the companies serving them.

Then there is the expense to site owners. Bots don't click ads, and can hit a single site thousands of times, which means the web's main revenue model is made obsolete by them while leaving someone to cover the bandwidth and servers. Cloudflare CEO Matthew Prince has floated charging bots a fraction of a penny each time they loaded a page. His example was an agent who asked where to eat lunch, which scanned every menu in town even though only one restaurant got the order.

Cloudflare sits in front of more than 20% of the web and was working with Coinbase and Stripe on a way to collect those payments. Who pays, who is liable and who is in control now remains to be seen.

Muse may have given a lot of people their first agent, but there is more to figure out before these tools go truly mainstream. The top spot has a way of changing hands every six months or so, and the next winner may be whoever figures these problems out first.