A caveat up front: what follows comes from internal documents cited by news outlets, primarily The Information, and has not been confirmed by Meta. That said, the picture emerging from multiple independent reports, including Forkast News and The American Bazaar, is consistent and worth paying attention to.
The project is called Hatch, and it would reportedly turn Meta from a social media company into a provider of autonomous digital agents able to navigate complex web environments like DoorDash, Etsy, Reddit, Yelp and Microsoft Outlook. Unlike traditional chatbots built to answer questions, Hatch is designed to carry out tasks on a user’s behalf: booking a table, comparing products, tracking a delivery, summarizing emails and calendar entries, all with the user’s authorization.
According to American Bazaar, the launch is expected within weeks, with the assistant operating as a conversational interface inside Instagram and WhatsApp. Users would describe what they want done and grant specific account access, a setup that immediately raises questions about how Meta will handle the data it collects, where it will be stored, and who oversees the actions an agent takes on its own.
On the model powering Hatch, both sources point to a codename, Watermelon, which internal documents claim has reached performance parity with GPT-5.5 on Meta’s own benchmarks, using roughly ten times the compute of its predecessor, Muse Spark. Worth stressing: these are internal, unverified claims, and the model itself isn’t expected to launch before October.
The pricing angle might be the most interesting part. Documents cited by Forkast suggest Meta is weighing a premium tier priced as high as $199.99 a month for heavy users, well above the existing Meta One Plus and Premium tiers already tested in markets like Singapore and Guatemala. The business side is already live: since August 1, the WhatsApp Business Agent has charged $2 per million tokens, roughly 4 to 5 cents per typical exchange.
There’s a technical wrinkle that makes the story more interesting: Hatch is reportedly inspired by OpenClaw, a viral open-source agentic tool built by Peter Steinberger that took off in November 2025. The irony is that Meta banned internal use of OpenClaw in February 2026 over security concerns, shortly before Steinberger himself left the company for OpenAI.
Mark Zuckerberg laid out the strategic logic during Meta’s July 29 earnings call, describing a three-part approach: keep the existing ad and recommendation systems running, scale business agents and APIs, and roll out consumer agents to the platform’s 3.5 billion users. He defended the company’s planned $125 billion to $145 billion in 2026 capital spending by arguing it would be foolish to basically just sell all of the compute and take a short-term profit.
Whether a price tag close to $200 a month finds real demand among consumers remains to be seen, and so does whether Meta can get people to trust an agent acting on their behalf inside apps that already hold private conversations and financial details. Until Meta itself confirms any of this, though, it remains a portrait assembled by reporters from internal documents rather than an official announcement, and it deserves to be read with that caveat in mind.
Meanwhile, Meta is already running smaller-scale trials of third-party agents inside WhatsApp, a step Forkast describes as laying the groundwork for Hatch itself. It’s the kind of incremental rollout the company has used before with other features, testing them in smaller markets first before a global release. Regulators in both the US and Europe have already shown interest in how AI agents handle account permissions and payment data, so any real launch is likely to draw scrutiny well beyond the tech press covering it today.

