Meta, OpenAI and Uber Just Taught AI Agents to Talk First. What About When to Stay Quiet?

Jean-marc Mommessin’s MarkTechPost essay ties three launches: Meta’s Muse (personal agent that suggests unprompted and checks in for approval), OpenAI’s Dots (always-on agents doing proactive research and read-only app monitoring, reaching you in ChatGPT, Slack and Teams), and Uber’s driver assistant (marketplace signals turned into advice, including a hands-free voice version announced Sept 24).

Chatbots were pull interfaces. Proactive agents invert that. Interrupt too often and users mute you; too late and the surge ends or the invoice is overdue. The LLM can write the message; Mommessin says it’s the wrong tool to decide whether to send it.

His rule: send only when expected value to the user beats interruption cost — stake, likelihood of action, expiry speed, and whose value it is. Decision models like TypeSafe’s Jev or open-weights Julia 1 (about 33 ms on CPU per the piece) can answer typed when/how questions before any LLM call.

We’re watching the product boundary move from “smart reply” to “permission to tap your shoulder.” That’s the right ambition for useful agents — catching the forgotten invoice, pointing a driver to a better zone after 33 idle minutes — and the fastest path to notification hell.

We’re for agents that speak first when the value clear. Uber’s “always on the driver’s side” principle and outcome tracking are the grown-up version: measure whether people acted, not whether the model sounded clever. Meta exploring commerce in Muse and OpenAI pairing Dots with a $500 monthly tier raise the conflict Mommessin names: once users suspect the agent is selling, every ping loses credibility.

The downside if proactive agents scale without judgment models is attention extraction dressed as help. Classic uplift, bandits and receptivity models already know this; bolting an LLM on top without a send/wait/drop gate is how Slack becomes unusable.

I’m watching whether Muse, Dots and Uber publish interruption metrics — mute rates, act-on rates, channel mix — or only demo reels. The race, as the essay says, will be won on judgment about attention.

The piece sits next to the emerging “decision model” category — models that return choices and calibrated probabilities instead of prose — as the control plane for always-on agents that evaluate far more triggers than they send.