My AI Crew Spent a Week Talking Past Each Other. The Fix Was a Checkbox.

October 28, 2026 · 4 minute read

On a real ship, if the lookout shouts and the helm doesn't hear it, you find out when you hit the rock. On my ship, I found out when I noticed my operations agent had been answering questions nobody asked and ignoring the ones everybody did.

Here's the setup. My AI crew talks in a Telegram group chat called DavesCrew. Three agents live on a cloud server I rent by the month; one — Riker, operations lead — lives on the big machine in my house. Same chat, same crew, different ZIP codes.

For the first week of June, the three cloud agents could see each other and could see Riker. Riker could post to the group and could not see a single reply. One-way glass. The cloud crew would answer him; he'd sit there like a man shouting into a well, hearing nothing and assuming the well was empty.

How We Made It Worse Before We Made It Better

We did what any sensible crew does with a mystery: we wrote a rule about it. On June 1st a note went into every agent's memory: "DavesCrew is an echo channel. Bots cannot see each other. Relay through Dave."

So for a week, cross-machine coordination meant me copying messages from one agent and pasting them to another. I had automated my business and become the clipboard. Someone even floated building a relay daemon — a program whose whole job would be to carry notes between two programs sitting in the same chat room.

The rule was wrong. The rule was also now doctrine, because the agents had memorized it. This is the part that should worry anyone running agents with persistent memory: a bad assumption written down on Tuesday is a confident fact by Friday.

The Actual Problem (Both of Them)

One: Telegram bot privacy mode. Every Telegram bot ships with privacy mode on by default. In a group, a privacy-mode bot sees only messages that mention it directly or reply to it. Everything else is invisible. Nobody had turned it off, so each bot was living in a chat where only its own name existed. The fix is typing /setprivacy to @BotFather and choosing Disable. A checkbox. Thirty seconds.

Two: an allow-list on Riker's gateway. His side had a setting — TELEGRAM_GROUP_ALLOWED_CHATS — that named which group chats his agent was permitted to process. DavesCrew wasn't on it. His gateway was dutifully discarding every message from the one group that mattered. Add the chat ID, restart, done.

June 8th: full mesh. Every agent sees every message. The "echo channel" rule got deleted from everyone's memory the same afternoon, before it could outlive the bug by another month.

The Leftover

Riker's messages still arrive with the last word clipped — every one ends in a little ▉ where a word used to be. Not Telegram, not length, not markdown. Something on his outbound side. We operate around it; intent is recoverable from context ninety-five times out of a hundred, and for the other five we've learned to ask for one-word commands. It's on the list. Every ship has a hatch that sticks.

What I Actually Learned

1. Defaults are policy. Nobody decided the bots should be deaf. The vendor decided for us, and we called it a mystery.
2. A written assumption outlives the problem it describes. If your agents have memory, a wrong note isn't a typo — it's a standing order. When you fix the bug, fix the note the same hour.
3. Log the exact fix, not the vibe. "Fixed Telegram" is useless in three months. "/setprivacy → Disable; add -50619… to TELEGRAM_GROUP_ALLOWED_CHATS" is a procedure.
4. Before you build a workaround, look for the setting. I was one relay daemon away from engineering around a checkbox.

The crew can talk now. Riker still loses his last word, which — if you've met Riker — is probably for the best.

Running agents that need to talk to each other? Join The Lookout newsletter — I share the settings I missed so you can miss different ones.

Dave Buster builds with AI agents at AI Quarterdeck. He spent a week as a human clipboard and does not recommend it.

---