Meta had a working version of its Muse AI agent by spring 2026, but waited roughly seven months to release it while building safety and reliability checks. The delay, described by Meta product chief Mona Sarantakos, offers a useful look at what it takes to move an agent from promising demo to something people can trust with everyday tasks.
Meta AI Watch analysis
What happened
Sarantakos told Lenny’s Podcast that Muse’s team spent the long middle phase of development on reliability, safety and security. Existing benchmarks did not fit an assistant that can use personal context and spend money, so the team built its own evaluations. She says each new capability brought a fresh attack surface: Muse launched without a browser because adding one raised security problems the team had not yet addressed.
The team also uses human-written system prompts and behavioural specifications, according to Sarantakos. BigGo Finance’s account of the interview says the agent’s evaluation work draws on real user conversations and failures, including cases where people get blocked while shopping.
Why it matters
An agent that can act across personal accounts needs a different standard of testing from a chatbot that mainly answers questions. The interview describes a team treating new capabilities as new risks to evaluate, rather than assuming that a successful demo means the product is ready. That is a concrete trade-off: fewer features at launch, with more time spent checking what the existing ones can do.
Our read
Seven months is a long time in the race to ship AI, and it is encouraging to hear a product leader explain what that time bought. The more useful test will be whether the evaluation work keeps pace as Muse gains capabilities. Agent safety is not a one-off launch checklist, however satisfying a tick-box may feel.
What to watch
- Whether Meta adds capabilities such as browser access, and what safeguards accompany them.
- How the team evaluates new tasks and responds to failures found in real use.
- Whether Meta shares more detail about its benchmarks and human review process.
Discussion spark: Should AI agents be held back until their makers can build task-specific safety evaluations, or should users help uncover the failures through earlier access?
Sources and evidence
- Meta's Muse Held Its Launch for Seven Months. Now It Has an Activation Problem – BigGo Finance (15 November 2026, 15:08 UTC)
not affiliated with, endorsed by, or operated by Meta