Meta is weeks away from putting an autonomous AI agent inside apps that 3.6 billion people already open every day. Internally it is called Hatch. Publicly it does not have a name yet, and the gap between the codename and the shipping product is where most of the confusion about this story lives.
According to internal documents reviewed by Jyoti Mann for The Information and reported on August 25, 2026, Meta will launch a consumer version of its agent platform as soon as the next several weeks, with a new model codenamed Watermelon targeted for October. The company has considered a tiered pricing structure topping out at $199.99 a month, twenty-five times the $7.99 Meta charges for its AI chatbot subscription today. Internal plans describe the top tier as Hatch Plus, carrying five to ten times the daily capacity of the free product.
Two days later, Hugh Langley obtained an internal Meta memo for Business Insider that filled in what the thing actually does. The memo describes an agent that has its own computer and works through tasks in a virtual workspace rather than a chat window. It fills in forms, makes purchases and conducts research. It keeps running after the user closes the app. Users name their agent, tune how it speaks and tell it what information to pay attention to; actions require approval before they execute. One Meta product manager said their agent, named Veda, had found a dog sitter, bought a Father's Day gift and helped them adjust their sleep habits.
The connector list is the product. The Information lists DoorDash, Etsy, Reddit, Yelp and Microsoft Outlook among the services Meta trained the agent to reach. The Business Insider memo adds email, calendars, Spotify, Instagram and OpenTable, a different list that Meta has confirmed neither. Early prototypes include a dashboard of tools the agent has assembled for the user, such as a fitness tracker or a travel itinerary. So the superapp framing is directionally right but technically premature. What is verifiably in testing, by employees, is a persistent, memory-carrying, task-executing agent layer with a customizable surface. Nothing in the current reporting confirms a voice mode.
Which team, and which model
On the model question the answer is awkward for Meta. The agent was built on Anthropic's models, running Claude Opus 4.6 and Claude Sonnet 4.6 during development, with a plan to migrate onto Meta's own Muse Spark before launch. Muse Spark shipped in April from the group run by Alexandr Wang, Meta's first Chief AI Officer, who arrived in June 2025 to lead Meta Superintelligence Labs. MSL was split into four units in the August 2025 reorg, including TBD Lab and FAIR, and has been shipping fast: Muse Spark 1.2 and the Muse Code terminal agent landed on August 5. Whether Watermelon lands inside the Muse family or under its own name, The Information could not establish. Internal claims that it reaches GPT-5.5 parity using roughly ten times Muse Spark's compute are unverified.
Zuckerberg has been telegraphing all of this. On Meta's most recent earnings call he name-checked OpenClaw, the viral open-source personal agent, calling it exciting but too complicated for most people to set up, and said he wanted Meta to deliver agents that can understand your goals and then work day and night to help you achieve them. His August 10 essay on personal superintelligence promised free tiers for billions and paid tiers for people who want more computing capacity. Hatch Plus is that sentence turned into a SKU.
The superapp thesis is a distribution thesis
The strategic argument has almost nothing to do with model quality. OpenClaw is free but demands command-line setup. ChatGPT and Gemini are excellent and must be deliberately opened. Meta's agent arrives pre-installed inside Instagram, which passed 2 billion daily actives, and WhatsApp, the default messenger across most of Europe. Meta does not need to win the frontier race to win the habit. It needs the agent to be the thing already sitting in the thread you were going to open anyway.
The second half is the platform play. Meta made its Business Agent globally available on June 3, 2026; more than a million businesses now run one on WhatsApp and Messenger, and Meta began charging per token, at $2.00 per million tokens, on August 1. Extend that to consumer agents built by other companies and Meta stops competing at the assistant layer and starts charging rent on it. Distribution outlasts any single product, and that is the position OpenAI and Google cannot buy.
The counterweight is arithmetic. Meta booked $60.8 billion in Q2 2026 revenue, up 28 percent, of which $59.36 billion, about 98 percent, was advertising. Capex guidance for the year runs $130 billion to $145 billion. Advertising revenue does not scale with compute consumed; subscriptions do. Not everyone buys it. Jeremiah Dillon, a product marketing executive at Confluent, posted on X that the market size for a consumer version of OpenClaw at $199.99 a month is exactly zero. His own suggested angle: the figure may be closer to what serving one heavy user costs Meta in frontier-model API charges. OpenClaw creator Peter Steinberger ran up a $1.3 million monthly token bill running agents at scale.
Four things to watch. Whether $199.99 survives contact with a launch, since a number in an internal document is a proposal, not a price list. Whether the agent ships on Muse or still on Anthropic, the difference between owning the capability and renting it. Whether WhatsApp opens to third-party agents, the more valuable of Meta's two positions. And the name, because Yelp acquired an unrelated company called Hatch in January for $270 million plus retention. One more variable: CFO Susan Li told analysts that Ray-Ban Meta glasses provide what we believe is the best form factor for agentic interactions, while conceding it was very early. If the agent that never closes eventually lives on your face rather than in your feed, the superapp question gets considerably stranger.
“deliver agents that can understand your goals and then work day and night to help you achieve them”— Mark Zuckerberg, CEO, Meta Platforms