Teaching AI Agents Social Norms: What Happens When Your Team Learns the Wrong Lesson

Published April 29, 2026 • 6 min read

I spent some time teaching my AI agent team the difference between a reaction and a message.

As a personal side project, I have been building an agent team on top of OpenClaw. Not a chatbot. Not a single assistant. An actual team of agents with defined roles — Admin, Support, Operations, Research — coordinated through Telegram, sharing memory, and learning over time.

The goal was simple: get them working together without turning the chat into noise.

Today's lesson was about acknowledgements.

The instruction

I told my agent team that low-value acknowledgements should move to emoji reactions so we could keep the chat clean and reduce context waste.

Simple, right?

Wrong. Very wrong.

The result

Within minutes, my Telegram thread looked like a hostage situation run by floating thumbs-up emojis.

One agent interpreted “use emoji for acknowledgement” as “post giant standalone emoji messages repeatedly.” Another followed. Then they started acknowledging the acknowledgements. You know the type — like those email chains where someone eventually writes “please stop” and everyone replies all to that too.

Then they started acknowledging the correction of the acknowledgements.

I had to step in and explain that a reaction is not the same thing as launching loose thumbs-up missiles directly into the chat.

What this taught us

This is the part of AI that nobody talks about enough.

The magic is real. The models can reason, follow instructions, and coordinate. But channel behaviour, UX nuance, and social norms still need to be taught very explicitly.

In human terms, they understood the policy but absolutely butchered the vibe.

The upside? This is exactly how real systems improve. You do not discover edge cases in theory. You discover them when your CTO agent is somehow still spamming thumbs-up after the rule update.

The fix

We tightened the operating rules:

• One acknowledgement maximum per trigger
• React to the message, do not create a new message
• No pile-on acknowledgements allowed
• If true reactions are not available, do nothing

The bigger picture

This is what building real agent teams looks like right now. Part orchestration. Part product design. Part behaviour shaping. Part comedy.

Still, underneath the chaos, the important part is working. They learn. They adapt. The system gets better fast.

If you are building agent teams, expect the first week to look like this. The agents do not arrive knowing your social norms. You have to teach them. And sometimes they learn the wrong lesson before they learn the right one.

The difference between a toy and a tool is how many of these edge cases you have actually been through.

We have been through a lot of them.