Three recent launches — Meta’s Muse, OpenAI’s Dots, and Uber’s driver assistant — share one bet: the agent speaks first. The hard problem shifts from what to answer to when to interrupt, on which channel, and with what offer. Miss the moment and the surge or invoice is gone; overdo it and users mute you. The good news is clear: LLMs write, but decisions need models. Classic ML and new decision models can judge value, timing, and channel with restraint.
Three launches, one pattern
The common thread is simple: the agent acts proactively, initiates contact, and runs continuously. It processes triggers in the background, returns with advice, and reaches you in the right place. When needed, it asks for approval and shows evidence. This is not pull chat; it is a service that watches context and suggests the next step.
Meta’s Muse (September 8) 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. You interact with it in its own app and in WhatsApp, receiving timely nudges during the day.
OpenAI’s Dots (September 29) are always-on agents for “proactive research.” They run read-only monitoring of your apps to catch a forgotten invoice or a bug in Slack. Dots reach you in ChatGPT, Slack, and Teams, with text and voice coming. You do not ask; they highlight risk and propose an action.
Uber’s driver assistant turns live marketplace signals into advice. In a recent talk, the team described a case: a driver idle for 33 minutes was pointed to a better zone, with evidence. Their principles are clear: stay always on the driver’s side, surface opportunities proactively, and measure whether drivers acted. A hands-free voice version was announced on September 24. Is this the new interaction standard?
From pull to push
Chatbots were a pull interface: you chose the moment, the channel, and the question. Proactive agents invert all of that. Interrupt too often and users mute you. Interrupt too late and the surge has ended or the invoice is overdue. Pick the wrong channel and a good message fails. The LLM can write the message; it is the wrong tool to decide whether to send it.
Attention becomes the currency. It is easy to spend and hard to regain. Every interruption must justify itself on the spot. Otherwise you lose trust, and with it the permission to speak next time. Shouldn’t we set a simple rule for all agents?
The main challenge is no longer what to answer. It is when to interrupt, on which channel, and with what offer. These three variables are coupled and can ruin experience if one is wrong. That is why send decisions need a separate judgment model, not another LLM pass.
This reframing moves the task to expected value, risk, and context. We no longer ask whether the agent can write. We ask whether it is worth attention, when to send it, and where.
The rule: value must beat interruption cost
Every proactive message is a bet. Send it only when its expected value to the user exceeds the cost of interrupting them. This single rule disciplines product and channel at once.
Value has four parts. How much is at stake right now. How likely the user is to act. How fast the opportunity expires. And whose value it is — the user’s or the platform’s. Only the balance of these parts answers the question honestly.
“Send only when expected value to the user exceeds the cost of interrupting them” — the core filter for a speak‑first agent.
Value then sets both decisions — timing and channel. High-value, expiring items go now. Modest ones wait for a digest. The rest are never sent. And the more a message is worth, the more intrusive a channel it has earned.
That creates a ladder: from in‑app card to chat, then to SMS, and finally to voice. Each step is more expensive in attention, and only justified when the user’s gain is unquestionable. Isn’t that the fair contract?
“When” and “how” are prediction problems
Deciding when and how is not a text task; it is prediction. Notification and growth teams have solved versions of this for years. They estimate lift, find moments, and judge whether a user can engage right now.
Uplift models estimate whether a nudge truly causes the action. They separate “it would have happened anyway” from “the agent made it happen.” In turn, contextual bandits learn the best moment and channel for each user with minimal live experimentation.
Receptivity models predict whether the person can engage right now: driving or parked, in a meeting or free. Combining these estimates yields a practical plan: who to reach, when, and in what voice.
Channels have a price of admission. An in‑app card costs little, chat costs more, SMS more still, and a voice call is reserved for urgent, hands‑busy moments. Uber’s outcome tracking supplies the labels these models need.
A channel is not just format. It is a bet on attention: card — cheap, chat — pricier, SMS — pricier still, call — for urgent.
Where decision models fit — and the limits
A new class of models suits this job. Decision models such as TypeSafe’s Jev and the open‑weights Julia 1 from Supersonic Labs return no text, only typed judgment: a choice from a short list, a score, or a yes/no, each with a probability. They are cheap, fast, and produce clear decisions before any LLM call.
Julia 1 runs on a CPU, from about 33 ms per decision. For an always‑on agent, that matters: it evaluates far more triggers than it sends. So each candidate first goes through cheap, structured questions.
Ask the simple things: is this worth an interruption, how valuable, and how urgent. Send now, wait, batch, or drop. And which channel: card, chat, SMS, or voice. This builds a disciplined decision gateway for any message.
The limits matter. These models judge only the context they are given, so they will not learn that one user ignores mornings. That demands careful signal design and humility in conclusions. And one more caution — cross‑sell.
Cross‑sell, carefully. An agent that speaks first is a powerful distribution channel. Meta says it is exploring commerce in Muse and has already launched Muse for Small Business; OpenAI announced Dots alongside a new $500 monthly tier; Uber has delivery and other services to promote. But an offer is just another message. It must clear the same bar, measured on value to the user. Once users suspect the agent is selling instead of serving, every message loses credibility. In the end, judgment about attention wins: speak 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.
Based on Meta, OpenAI and Uber Just Taught AI Agents to Talk First. What About When to Stay Quiet?.
