Imagine hiring a personal assistant who has been quietly reading your mail since 2011. They know which ex you still check up on, which sneakers you almost bought at 1am, which parenting group you lurk in without posting. They are extremely good at their job, precisely because of all that. Now imagine you are supposed to feel excited about handing them your to-do list.
That is roughly the pitch behind Hatch, Meta’s AI agent platform, which the company plans to launch in the coming weeks. It will take instructions and carry out tasks on your behalf. It will start behind a waitlist. And it will run on social data from Instagram and Facebook.
That last detail is the whole story. Everything else is table stakes.
Why the data angle is the product
Every agent product shipping right now has the same cold-start problem: the model does not know you. You have to explain your preferences, your constraints, your budget, your tolerance for a two-hour delivery window. The setup tax is real, and it is why a lot of people try an agent once, sigh, and go back to doing the thing themselves.
Meta does not have that problem. It has over a decade of behavioral signal on a couple of billion people. Where you live, who you talk to, what you scroll past, what you linger on. Feed that into an agent and the personalization is immediate in a way OpenAI and Anthropic cannot match without asking you a lot of questions first.
So Meta’s competitive advantage here is not the model. It is the file.
Which is also the reason to be careful. There is a meaningful difference between a company using your behavior to decide which ad you see and a company using your behavior to decide which purchase to make on your behalf. The first is annoying. The second is consequential. An agent that acts in the world and is informed by your social graph is a different kind of object than a chatbot, and it deserves a different level of scrutiny.
The questions I would want answered before joining the waitlist
Meta has not shipped this yet, so most of what matters is still unknown. Here is what I will be looking for on day one:
- Can you turn the social data off? If personalization from Instagram and Facebook is mandatory, that is a design choice, not a technical necessity.
- What is the audit trail? When an agent takes an action you did not expect, you need to see what it did, when, and on what basis. No log, no trust.
- Where are the spending limits? Any agent that can transact needs hard ceilings the user sets, not defaults buried three menus deep.
- Does data flow back? If the agent learns from your tasks and that learning improves ad targeting, say so plainly.
- What happens when it is wrong? Not “sorry, try again.” Who eats the cost of a bad order, a wrong message, a canceled reservation.
These are not hostile questions. They are the questions that separate a tool from a liability, and they are the ones I ask of every agent product that crosses my desk.
The waitlist is a tell, and not necessarily a bad one
Starting behind a waitlist means one of two things. Either the thing is not ready and Meta needs a controlled rollout to catch failures before they scale, or the waitlist is a demand-signal exercise dressed up as caution. Both happen. Both are common.
I lean toward the first reading, and I think that is the correct call. Agents fail in embarrassing, expensive ways, and doing that in front of billions of users at once would be a bad afternoon for everyone involved. A gated launch is the responsible version of this. It also means early reviews, including mine, will be based on a deliberately narrow slice of what Hatch can do.
My honest read
Meta is late to the agent race and arriving with the deepest personal dataset in consumer tech. That combination could produce the most genuinely useful assistant on the market, or the most uncomfortable one, and the difference will come down to controls and defaults rather than model quality.
If Hatch ships with clear opt-outs, visible logs, and spending limits you set yourself, it has a real shot at being the first agent that works out of the box because it already knows the context. If it ships with personalization welded on and the data flow unexplained, it will be a very capable tool that I would still hesitate to recommend.
We will find out in a few weeks. I will be on the waitlist, notebook open, checking the settings page before I check the features.
🕒 Published: