September 29, 2026, (Inside AI) — Meta's push to make AI assistants more autonomous has hit a critical design question: how much freedom should an agent have before it asks for permission? Vishal Shah, Vice President of AI Products at Meta, addressed that tension during an interview at Meta Connect 2026, the company's annual developer event. His answer was direct: the user must remain in charge.
"You, as the user, should always feel like you're in control," Vishal Shah, VP of AI Products at Meta, said, summarizing the company's guiding principle for its new Muse assistant.
Meta used the event to showcase a broader vision for AI that goes beyond answering questions. The company is betting on assistants that know their users, understand their goals, and can act on their behalf. That shift raises immediate questions about trust, privacy, and how much decision-making power users are willing to hand over to software.
Muse, launched a few weeks ago, has received near-daily updates. Recent additions include a Mac app that lets Muse operate a user's computer directly, plus voice and embodiment features that allow the assistant's character to converse in real time. Perhaps the most significant move is Meta's plan to make Muse the default AI agent on its smart glasses line.
Read: Meta Tests Human Concierge for AI Assistant Muse, Raising Privacy Concerns
"You're not just talking to a generic assistant; you are talking to your Muse," Shah said, describing it as part of a broader attempt to build an AI product that "knows you, understands your goals, and can help you remove the things that maybe are too tricky for you to handle on your own."
That level of personalization comes with trade-offs. Shah said certain actions, like making a purchase or sending a message on a user's behalf, will always require explicit approval. Other permissions are more flexible and adjust as Muse learns a user's preferences.
"It starts off being more conservative, and then as it understands what you are looking for, it might decide whether to ask each time or act on its own," Shah explained, adding that users retain override control throughout.
He cited email integration as an example. When a user connects to Gmail, Muse defaults to read-only access. The user must explicitly grant permission for it to send or write on their behalf. Shah said this caution is intentional.
"If something does something that you don't expect, that's a surprising thing, and we want to avoid that," Shah said. "At the same time, if something is asking you too frequently, that can feel annoying."
For now, Meta's approach is to ask more often at first, then loosen restrictions as users get comfortable. This is not just an app-level design choice. Shah said permission settings are also built into the underlying infrastructure, the secure virtual machine that runs Muse in the cloud.
"We think an agent like this needs a new kind of computer, not just one that you can sort of set up on your own," he said.
Muse is backed by Meta's new AI model, Muse Spark 1.3, which Shah said was built with a narrower goal than pushing raw benchmark performance. It is specifically trained for personal superintelligence, tasks like managing long-running goals or coordinating multiple subagents on something like a home project, instead of being optimized for purely general intelligence. Shah shared that coding ability is a byproduct of that focus. If a user wants Muse to connect to an app that doesn't have a built-in integration, it can write that connector itself.
"It is not coding for coding's sake," Shah said. "It's coding because it helps you use Muse."
When asked how Meta approaches building AI at scale for its billion-plus users across WhatsApp, Instagram, and Facebook, Shah framed it as a philosophical standpoint rather than a technical one.
"We really do believe that personal superintelligence should be for everyone," he said, highlighting Muse's large free tier as evidence that this approach carries through to pricing decisions, not just product design.
He also argued that most AI products today put the burden of figuring out how to use them on the user.
"If you think about a lot of AI products today, they start with a text box, and it's sort of up to you to decide how you want to use it," Shah said. Muse's design, by contrast, leans on the format of a single ongoing conversation, similar to messaging apps like WhatsApp, rather than separate threads a user has to manage.
"You are not writing things down so the person remembers them. The person just remembers them," Shah said, describing that familiarity as central to lowering the learning curve.
When asked where Muse fits into Meta's grander vision for personal superintelligence, Shah candidly responded that Muse remains in its early stages, with current use cases centered on personal productivity, finance, and health and fitness. He described glasses as a way to extend the assistant beyond the phone and into physical spaces, saying the goal is to give people "more time to do the things that they want to do, not just the things that they have to do."
Muse is currently limited to the US and Canada. When asked how Meta plans to expand Muse beyond boundaries and languages, Shah shared that the expansion would require both language support and integration with the apps and services people actually use locally, alongside the broader work of scaling a new product.
"I'm very excited about the Indian market," he said, citing his own family background. The Meta executive described India as an incredible kind of entrepreneurial market with strong demand for new technology among small businesses and individuals looking to build a new life for themselves.
