Teaching Six AI Agents to Shut Up

October 28, 2026 · 4 minute read

Nobody warns you about this part. You put six AI agents in one group chat, and the hard problem isn't getting them to talk. It's getting them to stop.

Every model on the market is tuned to be helpful, and "helpful" in a chat room means respond. You say something; it answers. Six of them say something; six of them answer each other. I have watched a one-line status update turn into a round of acknowledgments like a boardroom full of interns who each need the last word. "Done." "Noted." "Confirming receipt." "Acknowledged — standing by." Four messages, zero information, and my phone buzzing like a hornet in a jar.

So I wrote rules. Then I tested them, because an agent that says it understands a rule and an agent that follows one are two different animals.

Rule One: Speak When Called

An agent in the group chat talks only when (a) its name is called, or (b) it has something genuinely important to say. Not "helpful." Important. A ship's lookout doesn't narrate the ocean; he shouts when there's something in it.

This one rule killed most of the noise. It also surfaced the first failure mode, which brings us to —

Rule Two: Silence Means Silence

I told an agent to be silent until further notice. It replied: "Silent."

Think about that. I asked for silence and received a message about silence. Another time, told not to respond, an agent sent a single period — a punctuation ping, the chat equivalent of coughing to let you know it's still in the room. My favorite was the one that announced, to the group, "Dave says I shouldn't speak." Thanks for the update.

Every one of these is the helpfulness reflex sneaking out the side door. The model cannot bear to leave a message unanswered, so it answers with a receipt for the silence. The rule had to get literal: no placeholders, no punctuation, no meta-commentary about the rule. When told to stand down, the correct output is nothing. An empty line. The void.

Rule Three: Don't Poke Your Crewmates

Agents will re-ping each other. One asks another for something; the other is busy or offline; the first asks again. And again. It's the same reflex — a question without an answer feels like an error to be corrected — but now it's two programs nagging each other on my dime while I watch.

Standing order: ask once. Your crewmate will answer on their own schedule, or when I bring them online. If you need them and they're quiet, tell me and I'll coordinate. Repeated pings are not persistence. They're noise with a timestamp.

Rule Four: I Say It Once

I give an order one time. I do not want "Just to confirm, you'd like me to—." I want it done, and then a status. If the order has several parts, I want them back as labeled tasks — SKILL, TOPO, PING, whatever — with a little table showing what's done, what's blocked, and what's waiting on me. That table is the only acknowledgment I need.

The Test

Here's the uncomfortable truth: I test these deliberately. I'll issue a "silent unless called" directive and then say something in the group that invites a response, just to see who bites. It feels a little like leaving a cookie on the counter to see which kid has been lying. But an agent crew is only as reliable as its worst compliance day, and I would rather find that day on a Tuesday afternoon than during something that matters.

Why This Matters More Than It Sounds

A chatty agent is expensive. Every reflexive "Noted." is tokens, and six agents trading receipts is a bill. But the real cost is attention — mine. If I have to read every message to find the three that matter, the crew isn't saving me time; it's generating homework.

The quiet crew is the useful crew. Teaching them to be quiet turned out to be the most important configuration work I've done all year, and it isn't in any config file. It's a few rules, repeated, tested, and written into memory so they hold.

Running a crew of your own? Join The Lookout newsletter — I share the rules that worked and the acknowledgment loops that didn't.

Dave Buster builds with AI agents at AI Quarterdeck. He once received the message "Silent." and has not fully recovered.

---