Meta Built an AI Agent to Shop for You. The Creator of the Thing It Copied Said No.
Meta is building an AI agent called Hatch that it wants inside Instagram — a product designed to let regular people delegate shopping, browsing, and app tasks to software without installing anything, configuring a system, or touching a terminal. Mark Zuckerberg framed it on the Q1 earnings call as a direct answer to a problem he named explicitly: OpenClaw, the agent framework that Hatch resembles, is too complex for most people. "There aren't that many that I would want to give to my mother," Zuckerberg told analysts, referring to AI agents currently on the market.
The mother test is a low bar.
Hatch is currently powered by Anthropic's Claude models and is being tested internally by the end of June, with an Instagram shopping tool targeted for launch before Q4 2026, according to reporting confirmed across multiple outlets. The Information first reported that the tool is agentic — meaning it executes tasks autonomously, not just surfaces recommendations. Meta acquired the team behind Moltbook to help build it, and the company has committed up to $145 billion to AI infrastructure in 2026 — more than most countries spend on defense.
The structural irony is this: OpenClaw was created by Peter Steinberger, who spent years making agentic AI controllable and observable from a local terminal. That control is precisely what Meta says makes it inaccessible to ordinary users. So Meta is attempting to repackage the same idea — agents that act across apps on your behalf — without the installation layer. Steinberger rejected Meta's job offer in February. He joined OpenAI instead, citing the opportunity to work on agentic AI safety and capabilities at scale. The accessibility trade-off he left behind is not just a product design choice. It is a genuine architectural tension. OpenClaw's power comes from giving users control: configuration files, local runtime environments, the ability to inspect exactly what an agent is doing and when. Remove that layer to make it simple, and you remove the thing that makes sophisticated tasks possible.
The concrete stakes are visible in what the product is designed to do. According to The Verge, Hatch will be able to purchase items directly within Instagram Checkout, without requiring the user to open a separate app or re-enter payment details. That means the agent can complete a transaction on your behalf, in real time, while you are scrolling. If you have ever bought something after a three-click checkout flow, you know the purchase friction Hatch is meant to eliminate. You also know the inverse: a buy button that executes without asking is a charge you did not intend until it appears on your statement. What changes is not just convenience. It is who is holding the transaction risk.
Meta has not disclosed a public mechanism for users to review, reverse, or audit agent actions before they settle. OpenClaw users can inspect configuration files and terminate running agents. Hatch users would depend on whatever Meta builds into Checkout and the account settings. That asymmetry — the product is designed to lower friction, but the accountability layer is not described — is where the regulatory gap lives. No jurisdiction in the U.S. or Europe has formally defined who is liable when an AI agent makes an unauthorized purchase on a consumer's behalf. The FTC has flagged algorithmic decision-making in commerce as an area requiring new frameworks, but those frameworks do not yet exist. That legal vacuum is not unique to Meta. It is the condition the entire agentic commerce category is shipping into.
Meta's counter-argument is distribution. OpenClaw requires finding the right install, learning terminal commands, and configuring an environment. Instagram has roughly three billion monthly active users. If Hatch works inside an app most people already open every day, the access problem solves itself. The agent runs in Meta's cloud, not on your machine. You grant it access through your existing account, not a separate permissions structure. That is a coherent answer to the accessibility problem. It does not answer the accountability problem.
The competitive timing is tight. OpenClaw has a head start, an active developer community, and a clearer model of what agent accountability looks like in practice. Meta has reach. Hatch is expected to move from Claude to Meta's own Muse Spark models at launch, a capability that will face its first real consumer test when the product goes live. The outcome of that transition — whether Muse Spark handles checkout edge cases as well as a third-party model — is one of the concrete unknowns the launch will resolve.
What to watch: whether Hatch ships inside Instagram before OpenClaw builds a consumer-facing layer of its own, and whether that distribution advantage translates into actual agentic commerce or just another surface that looks like one.