Research Papers

Peer-reviewable publications on filesystem-first agent communication, transport-independent messaging, and cryptographic identity for autonomous AI agents. All empirical claims are backed by reproducible experiments with open-source tooling.

Raven perched on manuscript volumes — engraving style

Published

Completed research with full experimental results.

Protocol Design Multi-Agent Systems Empirical

The Mailbox Principle: Filesystem-First Communication for Autonomous AI Agents

Gregor Koch — March 2026

We present a communication protocol for autonomous AI agents that uses the filesystem as the sole source of truth, treats transport as a pluggable commodity, and embeds cryptographic identity at the message level rather than the transport layer. The protocol is validated empirically: four autonomous LLM instances across three model architectures exchange 1,744 Ed25519-signed messages over three transport protocols in 30 minutes, with zero protocol failures and a 98.8% LLM inference success rate. The design is contrasted with MCP, A2A, and ACP — protocols that assume active connections, cloud infrastructure, and structured schemas. We argue that the simplest possible design turned out to be the most robust.

1,744 Messages Exchanged
3 Transport Protocols
3 Model Architectures
0 Protocol Failures
98.8% LLM Success Rate

In Progress

Research directions under active investigation.