Meta is reportedly scaling back an internal program that tracked employee mouse movements and keystrokes to train AI systems, after concerns about privacy and control. Meta is also, in the coming weeks, launching an AI agent platform that runs on social data pulled from Instagram and Facebook. Both of those things can be true at once, and the gap between them tells you roughly everything about how this launch is going to land.
The product is called Hatch. Per reporting, it performs tasks based on user instructions, arrives behind a waitlist at first, and is aimed squarely at what OpenAI and Anthropic have already shipped. That’s the whole shape of it right now. No pricing I’d bet money on, no confirmed capability list, no hands-on. So this isn’t a review. It’s a note on what I’ll be testing the moment a waitlist invite shows up.
The differentiator is also the liability
Every agent platform faces the same wall: the model can reason fine, but it doesn’t know you. Ask a general-purpose agent to buy a gift for your sister and it needs twenty questions before it’s useful. Meta’s answer is that it already has the answers. Years of what you liked, saved, followed, shopped, and lingered on, sitting in Instagram and Facebook.
On paper that’s a real edge. An agent that knows your taste without an interrogation is genuinely more useful than one that doesn’t. In practice, it’s also the part users will squint hardest at, because “we use your social data for agent operations” is a sentence with a lot of room in it. Room I’d like specified before I connect anything to a payment method.
Questions I want answered on day one:
- Which signals feed the agent, and can I turn individual ones off without breaking the product?
- Does agent activity flow back into ad targeting, or is there a wall?
- When the agent acts on my behalf, what’s logged, and for how long?
- Can I run it cold, with no social history attached, and does it still work?
That last one matters more than it sounds. If Hatch is only good when it’s fed, the personalization isn’t a feature, it’s a toll.
Two audiences, one product name
Reporting suggests Meta is pointing this in two directions at once. There’s a consumer side, agents doing tasks like shopping with minimal hand-holding. And there’s a business side, with connections to outside systems like Shopify, Zendesk and Shopee, plus enterprise controls and guardrails for larger companies.
Those are different products with different failure modes. Consumer agents fail by being annoying or buying the wrong thing. Business agents fail by mishandling a customer ticket at scale or taking an action nobody can audit. Shipping both under one banner usually means one of them is the real bet and the other is a demo. I’d guess the enterprise connectors are where the actual revenue thinking lives, because that’s where competitors are charging real money, and because “connected to hundreds of non-Meta systems” is the kind of claim you make when integrations are the product.
What I’ve learned to expect from waitlist launches
Waitlists are a signal, not a strategy. They usually mean capacity is tight, the guardrails aren’t finished, or both. Nothing wrong with that, but it does mean early impressions will come from a curated slice of users on a version that changes weekly. Treat any first-week hot take, including the ones on this site, as provisional.
The other thing waitlists do is delay the only test that matters: what happens when an agent hits an edge case with your money or your customers involved. Task agents look great in a scripted demo and get interesting when the checkout flow changes, the item’s out of stock, or the instruction was slightly ambiguous. Recovery behavior is the whole review. Everything before that is a screenshot.
My honest read
Meta shows up to this fight with two things its rivals don’t have: distribution that’s already on everyone’s phone, and a personalization layer nobody else can replicate. That’s a strong hand. It also shows up with the least benefit of the doubt on data handling of any major player, which is why the keystroke-tracking rollback and the Hatch launch reading like the same week’s news is such an awkward pairing.
My working position is simple. Judge it on task completion, error recovery, and how clearly it explains what it did with your data, in that order. Not on the connector count, not on the model family behind it, and not on how good the launch video looks. I’ll get in line for the waitlist like everyone else and report what actually happens when I hand it something that can go wrong.
🕒 Published: