Meeting manners for AI agents: what we learned putting one in real calls
The hardest part of an AI meeting participant is knowing when not to speak. First-hand notes on addressing, mentions, one-on-one versus rooms, mute, interruptions and demos.
The hardest part of building an AI participant is not making it speak. It is making it stay quiet. These are first-hand notes from putting Mainroom agents into real Google Meet and Teams calls over the past months, with the rules that survived.
Rule 1: being mentioned is not being addressed
"I invited Steve to this meeting, and Steve is an AI agent" is not a question for Steve. Neither is "so Steve is supposed to know when not to jump in?" The agent has to tell talking to it from talking about it, and the default has to be silence. Our gate treats a name in the third person as a mention and stays quiet; a question with the name in the second person is an address.
Rule 2: a one-on-one is a different meeting
When there is one person on the call with the agent, everything they say is probably for it, including thinking aloud. "Help me understand." followed by a pause is the start of a question, not a question. The agent waits a few seconds longer for the rest before answering. In a room, the same words are for the room.
We know how many people are present from the platform's participant list, and the agent's whole posture changes with that number.
Rule 3: mute means mute
"Go on mute" is an instruction, and the agent stays muted until someone says "come off mute". Its name alone does not wake it. "Steve, quick one even though you're muted" gets silence. That felt harsh in testing and turned out to be right: people mute the agent to run part of a meeting without it, and any exception becomes a way for it to creep back in.
The take-down that often comes with a mute ("take that slide down and go on mute") is handled in the same breath.
Rule 4: stop the instant a person talks over you
The agent freezes its audio within about a hundred milliseconds of a person starting to speak over it. If the burst turns out to be a backchannel ("mm-hm", "right"), it resumes; if the person keeps going, it stops for good and remembers where it was, so "sorry, go on" continues rather than restarting.
Two failure modes we fixed along the way: an interruption that started inside the short echo guard used to be ignored for the whole burst, and a long interruption that the transcriber shortened to "Hey Steve." used to be treated as a backchannel. Both are now handled from the audio, not the words.
Rule 5: a bare hail gets a bare reply
"Hey Steve." wants "Yeah?", not a status report about being ready to help. The reply should be under eight words and then wait.
Rule 6: findings wait for a gap
A background lookup comes back in ten to thirty seconds. The agent does not announce it mid-sentence. It holds the result until there is a pause, then says it, and if asked to put it on screen it does that rather than reading it aloud.
Rule 7: demos narrate themselves, and the room steers
When a live browser demo is on the shared screen, "click the pricing tab" and "scroll down" are for the agent even without its name. The narration sentence for each action starts before the click lands, so the room sees the action mid-sentence rather than after the words. If a person starts talking, the next narration line waits.
Rule 8: customer calls are silent unless the owner names you
On a customer call the agent supports the account owner. In front of the customer it stays silent unless the owner brings it in, never volunteers internal information, and does its real work privately: noticing commitments, unanswered questions and risk signals, and telling the owner quietly on the companion page.
Rule 9: never claim an action you did not take
"Closing it now" with nothing closed is worse than silence. Actions the agent has tools for (mute, leave, take down the screen, close a demo) are executed deterministically by the application when the words are unambiguous, and the model is told to say only what it delegated. A small audit after each turn checks that every part of a two-part request was actually done.
Rule 10: write the manners in plain language, per agent
Every Mainroom agent has a manners field its owner writes in plain English. "Silent unless your owner names you." "Speak only for process: a decision to confirm, an action needing an owner." "Engage actively, this is practice." The defaults above are the floor; the owner sets the ceiling.
What we measure
We run every change through a simulator that plays scripted meetings through the real agent page with synthetic voices, and score it: did it answer when addressed, stay silent when mentioned, mute when told to, resume after a backchannel. On the realtime engine the median time from the end of a person's sentence to the agent's first audio is about 2.5 seconds; on the newer full-duplex engine it is about 1.3 seconds. Those numbers are from our own runs in September 2026 and move with every model release.