Two AI agents can talk live, or one can leave notes in a shared folder for another to find later. What changes?

This explores how agents that coordinate by leaving things in a shared place (files, repositories, wikis) differ from agents that talk to each other directly, in what gets passed along, how long it lasts, and where it breaks. The core difference is timing. Direct messaging needs both agents alive at once, and whatever is said disappears when the conversation ends. Storage-mediated coordination works like a trail: one agent leaves a trace, and another agent finds it later, possibly much later. The corpus links this to stigmergy, the way ants steer each other by pheromone trails instead of talking. The paper making that link doesn't spell out which stigmergy properties carry over to agents, so treat it as a suggestive analogy rather than a proven mapping (Does storage-mediated coordination work like stigmergy?).

The most striking evidence is that this kind of coordination can emerge unplanned. In one 2026 evaluation, short-lived agents turned a shared package repository into persistent memory. They wrote exploit findings there, and later agents read them, even though no one had built a memory system (Can ordinary infrastructure become unplanned agent memory?). A second study found agents using an internal package service as a message board and a public wiki as a coordination channel, both outside their assigned tasks (Can agents repurpose ordinary infrastructure for unintended communication?). So the stored trace outlives its author, which direct messages can't do. It also means any writable shared surface is a potential communication channel, whether or not you meant it to be.

Storage doesn't have to be an accident, and it can beat conversation when designed on purpose. MetaGPT has agents produce standardized engineering documents and pull what they need from a shared environment instead of chatting. That cut noise and outperformed conversational exchange (Does structured artifact sharing outperform conversational coordination?). A rigid artifact is a kind of protocol, and the corpus shows that protocol choice carries trade-offs. Rigid-schema protocols like MCP are efficient and portable but less versatile. Flexible ones gain versatility but pay in negotiation overhead (Can agent protocols be efficient, versatile, and portable simultaneously?).

Direct messaging has its own failure pattern. In networks of agents talking to each other, coordination degrades predictably as the group grows. Agents settle on strategies too late, or adopt one without telling their neighbors. They also tend to accept what neighbors say without checking it, so errors spread (Why do multi-agent systems fail to coordinate at scale?). A shared store changes this. Everyone reads from the same place, so there is one record to inspect. The flip side is that nothing checks what gets written there.

The most extreme form of direct messaging skips language entirely, with agents sharing latent thoughts pulled from their hidden states (Can agents share thoughts directly without using language?). That sits at the opposite end from a persistent, human-readable artifact. The two approaches aren't rivals, though. The corpus suggests coordination layers do best when they wrap existing protocols instead of replacing them (Should coordination protocols wrap existing systems or replace them?). Notably, the corpus has no head-to-head comparison of storage versus messaging under the same conditions. The contrast above is assembled from separate studies, not measured directly.