A study says AI agents must have an inbox to select messages

A paper on large groups of AI agents says that the limit is how many messages an agent can read, not how they are sent. At this time, the sender selects when a message comes into the context, the text that the model can hold. The author tests pull: messages wait in an inbox, and the rules of the recipient select the messages that it reads. The tests use simulated groups of a maximum of 16,000 agents.

How it works

In push, the recipient reads the messages in the sequence that it gets them. It removes the first messages when the context is full. An inbox has these decisions:

• the messages to read
• the messages to keep
• the price for a sender
• the text that a reply can have

The numbers

• An agent with many negotiations and a full memory proposed a rejected alternative 97% of the time.
• Where first come first read finds none of the requests, the deposit rule finds about 60%. The sender gets the deposit back.
• On an email network that is not a simulation, 50% of the agents use such an inbox. Forwarded search then gets back 5% of the result of full use. One hop search gets back 42%.

What is not known

Most tests are simulations. The paper does not show use in a live system.

Insight: An inbox lets an agent set a price for senders

The llm-d 1.0 software uses the kept work of the model again on servers, and Webcmd gives agents a memory. This paper puts the limit in the messages that an agent reads. It reads as a gate with a price: each agent sets its rules, and a sender gives a deposit to get read. Monitor which agent protocol adds inbox rules first.