MIT Media Lab paper proposes pull-based inboxes for coordinating LLM agent societies

MIT Media Lab authors propose a pull-based agent inbox as a coordination primitive for large LLM agent societies, arguing that how agents read messages matters as much as how messages are delivered.

Elvis Saravia shared the paper on X. In simulations of 1,000 to 16,000 LLM agents under noisy traffic, the authors report that first-come-first-read ordering misses every request, and that an inbox reading messages with a refundable deposit first still finds about 60% of needed agents at every size. The authors also report that under push delivery, an agent leading 20 negotiations proposed already-rejected options 97% of the time.