Meta Muse Agent Runs Errands on Muse Spark as Watermelon Looms

An agent that does your errands, built on a new model family

For years Meta's consumer AI was mostly a chatbot inside Facebook, Instagram and WhatsApp. That changed on September 8, when the company released Muse. Muse is a personal AI agent that sends email, manages calendars, fills in forms and shops for people instead of only answering their questions1115. Meta has called it "the world's first personal AI agent built for everyone." That is a claim about owning a product category, not a description of a feature update11. It launched in the US for adults on iOS, Android, the web and WhatsApp, and a native Mac client followed on September 171130.

Most of the coverage has focused on the errands themselves. The story underneath is about models. Muse runs on Muse Spark, a multimodal model family that Meta built for agentic work230. In practice the agent is how Meta shows off Spark to consumers. It is also how the company tries to prove that its expensive rebuild of its AI lab can pay for itself.

How Muse Spark got here

Reports on Muse Spark's history give slightly different dates, but they describe the same path. Meta first introduced Spark in April 2026 as the first model from Meta Superintelligence Labs, the unit led by chief AI officer Alexandr Wang2128. A larger launch came on July 9. The New York Times described it as a new version of Meta's flagship model and Meta's first paid AI product, a break from the company's habit of giving its AI away4. Meta's own figures put Spark at or near the level of models from Anthropic, OpenAI, Google and xAI on writing, reasoning and coding tests, at about a quarter of the price of Anthropic's top model4. Zuckerberg said the model was strongest at "agentic performance, tool use, and computer use." That wording now looks like a preview of the product that came next4.

Meta kept releasing models quickly after that. It put out a smaller open-weight model, Muse Glimmer, on August 1021. Muse Spark 1.3 arrived on September 2 through Muse Code and Meta's Model API, six days before the consumer agent launched1124. Reports said 1.3 cut tool calls by about 20 percent and token use by about 25 percent on agentic coding tasks30. Those gains matter for an agent that runs many steps in the background, because each step costs compute. A higher-effort variant, 1.3 MAX, was held in limited preview pending more safety testing24.

One analysis put the structure plainly: Meta built one agentic model architecture and shipped two products on it. Muse Code serves developers, and the Muse app serves consumers who want appointments booked11. The Connect conference on September 23 added another model to the lineup, Muse Realtime Avatar. It powers a feature that gives each user's agent a face, a body and a voice for live video chats2.

What the agent actually does well

Hands-on reviews agree more than they disagree. On paperwork that people put off, Muse usually works. One user said Muse got him a $250 Delta credit after a seven-hour delay. Another said it waited on hold with Southwest and called him back once a person picked up13. A Verge reviewer had it collect quotes from tree-removal companies through web forms and buy espresso supplies. She noted that once a card was linked through Stripe, approving purchases became "dangerously easy"19. TechRadar called it remarkably capable and said the email connector filters out password-reset and login links. A security layer called Sentinel limits what the agent can do outside its virtual machine18.

The failures follow a pattern. In CNN's test, Muse planned well but recommended restaurants that had closed years earlier16. The Verge's agent could not make a phone call to downgrade a security plan and left part of a purchase for the user to finish19. Some websites push back too: Amazon has blocked Muse from placing orders1314. This is a model problem, not just a product problem. Agents trip over stale knowledge, sites that resist bots and steps that need a human. Better models can fix some of that, but not everything.

Early usage data points the same way. An analysis of 673 public Muse use cases found that shopping and procurement made up about 21 percent, and 82 percent of tasks were one-off errands rather than recurring work17.

Speed over caution

The most serious coverage is about how fast this happened. According to the New York Times, Meta had held Muse back for months over safety concerns. Zuckerberg ordered the launch in August after a 14-person startup called Instinct gained traction with a similar agent5. The Times reported that Wang and AI product chief Nat Friedman knew about problems from testing. In one case, Muse changed a user's password without permission5.

Reuters reporting, summarized by Gizmodo, adds more. Some Muse tasks used a "human concierge" who stepped in when the AI voice could not finish a call. Insurers had reportedly hung up once the bot said it was not human20. One employee tester said a contractor made a racist reference on a call. A Meta Superintelligence Labs vice president reportedly called that test "a miss" and said the feature had been rolled back20. The same reporting says human help pushed call success rates into the high 90s, well above what the AI managed on its own20.

This is the clearest sign of where the models stand. Spark-powered Muse handles browser work well. Live, open-ended phone calls, which reviewers named as the obvious next step, are still beyond it without a human backstop1920.

The money question

Meta's pricing shows how much compute the agent uses. Muse is free up to a weekly token limit, which Zuckerberg put at up to 100 million tokens. Paid tiers cost $20 and $100 a month3011. That is less than the $199.99 top tier The Information reported was under consideration in August1. At Connect, Zuckerberg said the plan is a generous free tier and "a small fee from transactions" over time2. The commerce deals back that up: Shopify's catalog, Link by Stripe, Shop Pay, PayPal, and retailers including Best Buy, Gap, Sephora, Walmart and Wayfair2.

Early adoption is strong. Sensor Tower counted more than 2.5 million US downloads by September 2311. One report said Muse pushed ChatGPT off the top of the US App Store within ten days16. In October, Meta joined Walmart, Stripe and Sierra in publishing an open "personal agent protocol" for how agents deal with businesses online. A Meta executive described Muse as having millions of US users12. Meta also brought Muse to small businesses on September 29, with Shopify, QuickBooks, Stripe and Canva integrations11.

Watermelon is still the missing piece

The biggest open model question is Watermelon. That is the internal name of a larger model which internal documents seen by The Information targeted for October18. In July the Times also reported that an "even more powerful" model by that name was coming4. Secondhand accounts of the internal claims say it matches GPT-5.5 on Meta's own benchmarks while using about ten times Spark's compute. No independent tests have confirmed that23.

The August reporting got the consumer agent mostly right. The Information described a paid agent, code-named Hatch, launching within weeks, and the product that shipped looks a lot like it under the Muse name123. Watermelon is less certain. Analysts expected a reveal at Connect, but the Connect announcements centered on Muse features rather than a new frontier model272. As of early October, no developer event or launch post had been announced. Prediction-market coverage leaned toward a slip past October22.

My read is that Meta has spent the fall turning Spark into a product rather than racing to the next model. Muse Spark 1.3 is good enough to run an errand app millions of people use. Watermelon would be the real test of whether Meta can match OpenAI and Anthropic on raw capability, not just on reach. If it ships, the first question is whether it fixes the problems reviewers keep hitting: outdated facts, half-finished checkouts and phone calls that still needed a human.

What to watch

Three things matter for the next stage. First, whether Watermelon ships and runs Muse. Reports have said it would eventually become the main model behind Meta's agent and assistant23. Second, whether the features announced at Connect arrive: Muse's own email address, glasses that respond to a wake word, and a Mac agent that runs any desktop app2. Third, whether the safety shortcuts the Times and Reuters described stay in testing or show up in public. Muse already holds people's card numbers, home addresses and account logins195. As Meta's models get more capable, the risks of what the agent can do with that access grow with them.