Agents are a group chat where the smartest member has the memory of a hamster
AI agents are a WhatsApp group chat where the smartest member has the short term memory of a hamster. To prove it, Aman Khan and I put on a puppet show:
(If you've been using group chats e.g. Slack or Whatsapp to get things done, you've been running AI agents for years. The only difference is now we are replacing some of the humans in the group chat with LLMs and tools.)
In this video, Laura Ackerman plays a human planning a picnic with the help of ChatGPT. I played the LLM, and Aman played a "weather tool."
Laura the user
Had to initiate and give the goal
Never calls the tool directly, can only ask Tal the LLM to call the tool
Tal the LLM
Compulsively responds to everything new message, whether tool or user.
Had to read the chat from the top every time (I was exhausted)
At some point reached context limit (Laura could have started a new group chat)
Aman the weather tool or "MCP server"
Needed specific input format
Doesn't pay attention to the chat, only when tagged
Whatsapp the "orchestrator" or "chat client" or "agent harness"
Whatsapp's main job is to host the growing text file that is the illusion of "chat."
Whatsapp is dumb: it just runs text and scans for @....... pattern and then sends a notification to Aman
A Whatsapp group is not even an analogy. It's literally what an AI agent is. To prove it, I exported the chat as a text file. Just like the Whatsapp chat is just a growing text file, an AI agent is also a growing text file that gets batted back and forth between OpenAI and the chat client.
Why this matters for building AI products
Once you see agents as group chats, the magic disappears and you can actually build.
The next step: move your workflows from ChatGPT or Claude to AI coding agents like Cursor or Claude Code. These tools show their work, so you can watch the reasoning unfold and see every tool call in real time. With the group chat analogy in your head, you'll be able to interpret what's happening instead of guessing.
The opposite of this would be purpose-built, polished UIs that abstract away all the tool calls and reasoning behind loading spinners and GIF animations. Great user experience, but not meant for building a strong sense of what these technologies actually are/are not.