In simulated societies of up to 16,000 agents, a push-based message protocol made one agent re-propose options it had already rejected 97 percent of the time
AspireAI community · arXiv
Agent infrastructure has optimised the wrong end of the problem. Protocols like MCP are good at delivering messages to agents; this paper is about what happens after delivery, where the scarce resource turns out to be reading, because every message costs tokens and context is finite. In the push model, the sender decides what enters the recipient's context, the agent reads in arrival order, and the oldest replies are dropped when the window fills.
The consequences are quantified. An agent leading 20 simultaneous negotiations with a full memory ended up proposing options it had already had rejected 97 percent of the time, simply because the rejections had been evicted. Under noisy traffic, an inbox that prioritises messages carrying a refundable deposit still finds about 60 percent of the agents a request needs, at every society size from 1,000 to 16,000. On a real email network the losses compound: when only half the agents adopt an inbox, forwarded search recovers just 5 percent of what full adoption would, against 42 percent for one-hop search.
The paper's conclusion is a design principle rather than a product: an inbox is a coordination primitive that each agent configures privately, and it decomposes into four separate decisions - what to read, what to keep, what a sender pays, and what a reply may say - each of which depends on the traffic pattern. The caveat is that the scale results are simulated with LLM-made decisions rather than drawn from a production deployment.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.