Meta appears to be building a scheduled tasks section for Meta AI on the web, where users could set up and manage recurring instructions instead of issuing every prompt by hand. Spotted in recent builds and not yet live for anyone, this feature would allow users to create jobs that run on a set cadence, such as a morning news digest, a weekly summary, or a periodic check on a topic, with the results delivered back inside the assistant. The naming leans toward the industry-standard "scheduled tasks," and the groundwork echoes an earlier Tasks effort seen some time ago, suggesting the next turn of that idea rather than a fresh start. What remains unclear is whether these jobs can trigger on events, reach connected apps, or push notifications when they finish.

The first beneficiaries would be web users who want Meta AI to act on its own, and Meta itself, which has spent much of this year closing the distance with rivals whose assistants already run scheduled tasks, ChatGPT among them. Recurring, unattended work moves the assistant from a question box toward something closer to an agent, the direction every major lab is pursuing.

The timing sits beside louder model news. Alexandr Wang, Meta's chief AI officer, told staff at a July 2 town hall that the company's next model, codenamed Watermelon and still in training, has caught up with OpenAI's GPT-5.5 on internal benchmarks and draws far more compute than Muse Spark, the April model known internally as Avocado.

Asked about matching Anthropic's Opus on coding, he pointed not to Watermelon but to a forthcoming Muse Spark update, due "pretty soon." No release date or public benchmarks exist yet, and Zuckerberg struck a more guarded note, citing a three- to six-month window for returns. How fast Watermelon ships will decide whether it meets today's frontier or arrives to face the next one.