Meta, OpenAI, Uber Push Proactive AI Agents

Proactive AI agents have quietly become the industry's favorite new idea. Over the span of three weeks, Meta, OpenAI and Uber each showed off systems that speak first: they watch, they decide something matters, they reach out. The shared bet is that the next wave of assistants will not wait to be asked. They will interrupt.

That shift sounds like a feature upgrade. It is closer to a change in what these products fundamentally are.

Three Launches, One Pattern

Meta introduced Muse on September 8. It is a personal agent that books, emails and keeps working with the app closed. It remembers details, makes unprompted suggestions and checks in for approval, both in its own app and inside WhatsApp. Meta says it is exploring commerce inside Muse, and Muse for Small Business has already launched.

OpenAI announced Dots on September 29. These are always-on agents that run what the company calls "proactive research," read-only monitoring of your apps, to catch a forgotten invoice or a bug in Slack. They reach users through ChatGPT, Slack and Teams, with text and voice coming. Dots arrived alongside a new $500 monthly tier.

Uber's driver assistant takes live marketplace signals and turns them into advice. In a recent talk, Uber's product team described a driver who had been idle for 33 minutes being pointed to a better zone, with evidence attached. The company's stated principles: stay "always on the driver's side," surface opportunities proactively, and measure whether drivers actually acted. A hands-free voice version was announced September 24.

From Pull to Push

Chatbots were a pull interface. The user chose the moment, the channel and the question. Proactive AI agents invert all three. The system picks the moment. The system picks the channel. The system decides whether there is a question at all.

That inversion moves the hard problem. Writing a message is now the easy part. Deciding whether to send it is not. Interrupt too often and users mute you. Interrupt too late and the surge has ended or the invoice is already overdue. And if you pick the wrong channel, a perfectly good message fails anyway, no matter how carefully it's written, no matter how urgent it feels, and no matter how much you've convinced yourself that this time, surely, they'll stop and read it.

A large language model can draft the note. It is the wrong tool to decide whether the note should exist.

Value Must Beat the Cost of Interruption

Every proactive message is a bet. The rule that follows is simple to state and hard to run: send only when the expected value to the user exceeds the cost of interrupting them. Value breaks into four parts. How much is at stake. How likely the user is to act. How fast the opportunity expires. And whose value it is, the user's or the platform's.

Value then sets both decisions. High-value, expiring messages go now. Modest ones wait for a digest. The rest are never sent. The more a message is worth, the more intrusive a channel it has earned. An in-app card costs little. Chat costs more. SMS more still. A voice call is reserved for urgent, hands-busy moments.

An agent that speaks first is a powerful distribution channel, but an offer is just another message. It must clear the same bar, measured on value to the user.

When and How Are Prediction Problems

Notification and growth teams have solved versions of this for years. Their tooling maps neatly onto proactive AI agents. Uplift models estimate whether a nudge actually causes the action, contextual bandits learn the best moment and channel for each user, and receptivity models predict whether someone can engage right now: driving or parked, in a meeting or free.

Uber's outcome tracking supplies the labels those models need. If the system knows whether a driver acted on the advice, it can learn which interruptions earned their place.

A New Class of Model for the Job

A different kind of model suits this layer. Decision models such as TypeSafe's Jev and the open-weights Julia 1 from Supersonic Labs return no text, only a typed judgment: a choice from a short list, a score, or a yes/no, each with a probability. Julia 1 runs on a CPU, from about 33 milliseconds per decision.

That matters because an always-on agent evaluates far more triggers than it ever sends. Each candidate can get cheap, structured questions before any LLM call:

• Is this worth an interruption?
• How valuable, how urgent?
• Send now, wait, batch or drop?
• Card, chat, SMS or voice?

These models judge only the context they're given. They won't learn on their own that one particular user ignores mornings, because nothing in the input tells them so, and nothing in their training lets them guess it either. Someone has to feed them that. And that's the whole problem.

The Cross-Sell Trap

Meta is exploring commerce in Muse. OpenAI launched Dots next to a $500 monthly tier. Uber has delivery and other services it would like to promote. The temptation to turn a proactive agent into a sales channel is obvious, and the risk is just as obvious.

Once users suspect the agent is selling instead of serving, every message loses credibility. The interruption budget is not infinite. Spend it on offers and you will have nothing left for the invoice that is about to go overdue.

Who Wins the Attention Race

All three companies have strong models. The race will be won on judgment about attention: speaking rarely, at the right moment, in the right place, on the user's side. Companies with years of notification data may have a head start, because the labels that train good judgment are the hardest part to assemble.

Proactive AI agents are not a smarter chatbot. They are a new interface with a new failure mode, and the failure mode is silence. Users will not complain. They will just stop listening.