Build1 distinct publisher3 min readPublished
Hatch is built to keep working with its app closed, so consent has to be banked up front and re-litigated later through an approvals surface. The pre-release build shows both, plus a recovery PIN that carries the private store between devices.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Consent for an agent that keeps working after you close the window has to be banked in advance, because there is nobody present to ask. That property, as Business Insider described it, decides the rest of the design [2]. The grant happens early, and something in the runtime has to be able to stop mid-task and wait for a human.
The pre-release materials reviewed by TestingCatalog show both halves. Onboarding runs three stages: connect accounts, choose starter tasks, design the agent's identity [3]. The screens are introduced with the lines "Hatch is better with connectors," "Put your agent to work," and "Customize your agent" [4], and the order matters: the connector grant is stage one, before the user has watched the agent do anything at all, which means everything they might later want to refuse depends on a surface they have not seen yet.
That approvals surface does exist in the build. There are dedicated experiences for voice, media, task progress, browser work, and approvals [5], and per TestingCatalog users appear able to see what Hatch is doing, review completed steps, and intervene before consequential actions [6]. The connector list is where this gets expensive: email, calendars, Spotify, Instagram, and OpenTable [7], against previously disclosed examples including filling out forms, making purchases, ordering food, and booking restaurants [8]. Purchases and bookings are writes that move money and do not get undone by closing a tab. If those queue up in a tray the way notifications do, the approval degrades into a receipt.
The privacy setup goes past an OS permission prompt. Hatch can prepare an encrypted private environment protected by a recovery PIN, with the PIN intended to unlock personal agent data on another device [9], and warning screens stop people continuing normally when the service cannot verify that their environment is protected [10]. The second part is the better piece of engineering. A hard stop rather than a dismissible banner is a fail-closed default, and it is worth naming when you see one.
The price signals who is expected to make that consent decision: Meta has considered tiers as high as $199.99 per month, according to reporting relayed by TestingCatalog [11]. Twelve of those is $2,399.88 a year [12]. That number only describes something real if the tier actually ships and if the capabilities in a pre-release build turn out to be the ones sitting behind it. TestingCatalog says plainly that some capabilities may remain restricted, change before launch, or arrive in later phases [13].
Treat the sample as controlled, because it is built that way. The service contains invite-code, waitlist, age-check, disclosure, and NDA flows [14], mobile onboarding references Apple TestFlight and the Google Play Store [15], and the iOS TestFlight build has already reached version 4 [16]. The people currently handing over a mail account at stage one signed up to hand over a mail account. Their tolerance is not a measurement of the general public's.
Ranked by verification strength, evidence, and original report placement.
Business Insider reported that Meta has expanded internal testing of Hatch beyond its Superintelligence Labs.
Business Insider described Hatch as a personal agent that continues working on a user's behalf even when its app is closed.
Hatch's onboarding process has three main stages: connecting accounts, choosing initial tasks, and designing the agent's identity.
Hatch introduces the three onboarding stages with the messages "Hatch is better with connectors," "Put your agent to work," and "Customize your agent."
The Hatch product contains dedicated experiences for voice, media, task progress, browser work, and approvals.
Users appear able to see what Hatch is doing, review completed steps, and intervene before consequential actions.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 28, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
invest
Meta's Hatch agent tops out at $199.99 a month, with DoorDash and Etsy behind the meter2 distinct publishers
product
Meta shipped Pocket the day Gizmo died, and never had to buy anything3 distinct publishers
product
Meta wants up to $199.99 a month for an agent, and it is selling a meter1 distinct publisher
product
Meta's 2026 Capex Plan More Than Doubles Two Years of Spending; the Product It Funds Is Still on Paper1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One private look, no second pair of eyes
Every specific detail here — the three onboarding prompts, the recovery PIN, the five task surfaces, the version-4 iOS build — comes from unreleased material TestingCatalog says it reviewed and no reader can see, filed under the byline 'Anonymous reporter'. Meta has not responded, and the parts that sound most confirmed, Business Insider's testing expansion and the $199.99 tier, are recaps rather than fresh checks. The internal consistency of the interface strings is what keeps this above rumour; the absence of anyone else looking at the same build is what keeps it well below established.
Employees and three build tracks
Nobody outside Meta uses Hatch. What can actually be counted is iteration rather than uptake: an iOS build on its fourth turn, Android and web tracks beside it, and testing said to have moved past Superintelligence Labs into the wider company. The invite codes, waitlist, age checks and NDA flows are machinery for keeping people out; reading them as traction would invert what they are for.
Interface strings read as strategy
TestingCatalog is honest about its limits and then reaches past them. Messenger Companion and WhatsApp Companion appear as references in an unreleased build and become a distribution advantage 'few competitors can match'; a set of surfaces becomes Meta's answer to superapps; September becomes a plausible launch. The $199.99 figure is somebody else's reporting, annualised to nearly $2,400 without a new check. The gap is moderate rather than severe precisely because the caveat about pre-release change is stated in the publisher's own voice instead of buried.
Scoop is the product
TestingCatalog's readership comes from finding things in unshipped builds, and this piece is labelled Exclusive and sourced to an unnamed reporter — a structure that rewards the most expansive plausible reading of what a screen implies. Meta's incentive runs the other way and is fully satisfied: no confirmation, no denial, and free anticipation for an unreleased product carrying a possible $199.99 tier. The countervailing signal is that the outlet marks its own uncertainty rather than dressing prototypes as features.
Confident on shape, thin on specifics
The broad picture is credible and coherent: a standalone agent that wants connectors early, keeps running with its app shut, shows its work, and gates consequential actions behind approvals with a portable encrypted store underneath. Confidence drops sharply on anything with a number or date on it — the price, the window, whether all three clients land together. Nothing in our coverage contradicts the account, but nothing corroborates it either, and pre-release screens are routinely cut before anyone sees them.