Today’s radar brings an adoption number for a consumer agent, research showing AI proving code security, and an open tool for keeping agents on a short leash. I’ll start with adoption.

When an AI product passes 10 million users in a few weeks, the conversation for people who work in product shifts from “will it work” to “how does it pay for itself.”

Meta said on its earnings call that more than 10 million people in the United States and Canada have set up Muse agents since the September launch. About a quarter use it every day, and the most common use is shopping, with an average ticket above 200 dollars.

The agent becomes a sales channel

What caught my attention most was the business model. Consumers use it for free and pay only for extra tokens or more advanced models, and the company plans to charge businesses when the agent helps complete a sale. In other words, the agent becomes a sales channel, not just an assistant.

This raises very practical questions for product people. If an agent buys on the user’s behalf, what happens to the journey we designed for the screen? How does a business show up, and get chosen, when another agent is the one deciding? Meta already talks about a protocol for consumer agents to talk to business agents, together with Shopify and Walmart. The guide on AI agents helps structure these decisions, and the one on AI agent operations covers the day to day.

The cost side

The cost side is also worth noting. Meta itself says Muse consumes a lot of computing capacity, which is why it started in two countries. In AI, scaling usage and scaling cost go together, and this pricing design has to be born with the product. The guide on AI agent cost management shows how to think about it.

I am optimistic about this path. Agents that solve tasks end to end tend to remove a lot of everyday friction. But those who build products will need to think about trust, spending limits, and how to be found by an agent, not just by a person.

For those who are curious and want the full numbers, here is the Fortune report.

The rest of the radar

  • Claude proves security of 123 libraries — formal verification done by agents may become routine in secure development, but “verified” is not “bug-free” and the cost is still high. Anthropic.
  • NVIDIA OpenShell — open runtime with a sandbox and credentials the agent never sees; if your product gives agents access, access policy becomes a requirement. GitHub.
  • The Jev hype — “valid schema” is not “correct answer”; measure quality on your own data and compare with open models before betting the roadmap. Alex Ewerlöf.

That’s it for today. More tomorrow.