Meta's Muse is the test for whether AI agents go mainstream
Greg and Colan Neese disagree on whether everyday people will hand an AI agent their email, and what happens when they do.
Meta's Muse showed up on billboards along the New Jersey Turnpike this weekend. Greg saw several on the drive home from Pennsylvania and took it as a signal: this is the consumer moment for AI agents. The open question is whether everyday people will give an agent what it needs to work. Greg and Colan Neese disagreed on the answer in a recording on 13 October, and the disagreement is the useful part.
What an agent does when you give it everything
Colan was already convinced. He gave Muse access to his email without hesitation, and he describes the payoff in small, practical tasks.
His example is dog grooming. He set a rule, every three weeks outside holidays. Muse goes to the groomer's website, finds times, checks the shared family calendar that is displayed in the kitchen, confirms the slot and sends the request. He just shows up with the dog. “I just gave it my email. That's all I did,” he said.
He also uses it for content. He has it ingest newsletters, podcasts and a back catalog, set up in Google Cloud so he can query across all of it, and compile a weekly re-brief for his team.
Why Colan calls it the worst agent you will ever use
Agents are not new, Colan said. They have existed for people willing to program one with tools like Claude Code or Codex. What Meta has done is package one for non-technical users, with a large amount of memory behind it.
“This is the worst agent you'll ever have.”
Greg had a parallel experience with ChatGPT, which he has used heavily since the start. When it pulled together what it knew about him, it understood his separate workflows, connected them, and surfaced something he owed someone from his email. He had forgotten it. That is the same pull Colan describes, from a different product.
The case against: the data barrier caps adoption
Greg's concern is that these tools are only as good as the data they get. Muse needs your email, and ideally more. His assumption is that a large share of people, not a fringe, would have to be willing to connect it, and he does not think they will be. His three-year view: agents keep improving for people in roles like theirs, creators and content teams, but do not reach everyday adoption. He compared it to VR, where mass adoption decides whether a technology survives. If Muse misses, he said, AI becomes a tool for specific groups of people.
The case for: privacy fears fade
Colan agreed that data access is the hang-up and argued it fades. He pointed to early internet users who refused to put a card number online and now keep their whole financial lives there. He worked in ad tech, and says personal data is already bought and sold widely, so a mailbox feels like a small marginal risk.
He also thinks current numbers undersell adoption. He disputes the line that only a sliver of people pay for AI, because it ignores employer-paid seats and free users. Most people's experience is the novelty version, asking the weather. The real shift, in his view, comes with Gen Alpha, who grow up with it the way millennials grew up with web interfaces and drove the rise of SaaS.
The security moment
Greg's sharpest point was not about privacy. It was about two-factor authentication. While watching Dots, his agent tool, work through a GitHub login, he saw it request a verification code, switch to his email, and pull the code. He also had a Google Messages tab open, so his texts were reachable too. The second factor he relied on was now readable by the same agent.
He extended the thought to banking. Nothing technically stops an agent from going to a bank's site, requesting a password reset, and reading the emailed code to complete it. Colan is not ready to connect bank or credit accounts himself, though he sees the value for financial planning.
Colan's three-year prediction: a serious incident, then a pullback
Colan expects things to get worse before they settle, and offered it as opinion. Open-weights models are close to widely available without the guardrails of the large labs. Much of the world's software runs on decades-old code and ancient APIs. An AI can work on that problem without tiring. His prediction is a serious incident within three years, a major cyberattack or something worse, such as damage to a power grid. The reaction will be regulation and a pullback, after which development continues. “Something pretty terrible is going to happen in the next three years with AI,” he said. Greg answered with a Cyberdyne joke and moved on.
What operators should take from it
The security point is the one to carry into your own work. Many Trust & Safety, support and account-recovery flows assume that email and SMS are separate channels. An agent with access to both makes that assumption weaker. Whether or not Muse reaches the mainstream, players and staff are about to use agents like it. Teams that design verification, recovery and fraud flows should ask now what those flows do when the person on the other end is an agent acting with the owner's permission.
Greg and Colan will pick this thread up on Never Entertained, on Fridays. For the weekly version, join The Operator's Briefing, and bring your own read to the Player Driven Discord.
