Meta, OpenAI, and Uber push agents that speak first; the hard problem becomes knowing when to stay quiet

Three separate launches in September reveal the same bet: the agent initiates. Meta's Muse, released September 8, runs in the background, pushes suggestions through WhatsApp and a dedicated app, and asks for confirmation. OpenAI's Dots, announced September 29, runs "proactive research" in read-only mode across Slack and other apps, then surfaces alerts in chat, Teams, and soon voice. Uber's driver assistant, unveiled by voice on September 24, translates real-time market signals into concrete recommendations — for example, a driver who has waited 33 minutes gets rerouted to a better zone with an explicit rationale. All three flip the interface from pull to push.

The problem: value versus interruption cost

Chatbots waited for the user — moment, channel, question. Proactive agents invert the equation. Interrupt too often and the user mutes; interrupt too late and the opportunity evaporates; choose the wrong channel and a good message dies. The large language model knows how to phrase the notification; it is the wrong tool to decide whether to send it. The guiding rule is simple: send only when the expected value to the user exceeds the cost of the interruption. That value breaks into four components: what is at stake, the probability the user acts, the urgency of the opportunity, and whose value is being served — the user's or the platform's.

Decision models, not text models

Notification and growth teams have been solving versions of this for years. Uplift models estimate whether a push causes action. Contextual bandits learn the optimal timing and channel per user. Receptivity models predict whether the user is free right now — driving or parked, in a meeting or at their desk. Channels carry an "entry price": an in-app card is cheap, chat is more expensive, SMS more still, and a voice call is reserved for urgent moments when hands are full. Uber's tracking of driver actions supplies the labeled outcomes these models need.

A new class: models that return structured judgment

Enter a young category — decision models. TypeSafe's Jev and Supersonic Labs' Julia 1 (open weights) do not generate text; they emit a typed judgment: a choice from a short list, a score, a yes/no, each with a probability. Julia 1 runs on CPU at roughly 33 milliseconds per decision. An always-on agent evaluates far more triggers than it ever sends, so every candidate passes through a cheap, structured filter before an LLM call: is this worth an interruption? How much value, how urgent? Send now, wait, bundle into a digest, or discard? Card, chat, SMS, or voice?

Limitations and commercial risk

The models judge only the context they receive. They will not learn on their own that a user ignores morning messages unless that signal is fed in. On the commercial side, an agent that speaks first is a powerful distribution channel. Meta is testing commerce inside Muse and has already launched a small-business version. OpenAI positioned Dots alongside a $500-per-month tier. Uber has deliveries and services to promote. But an offer is just another message, and it must clear the same user-value bar. The moment the user suspects the agent is selling instead of serving, the credibility of every message collapses.

The bottom line

All three companies have strong models. The race will be won on attention judgment: speak rarely, at the right moment, in the right place, through the right channel.