Self-Evolving Agent Communication Protocols
A Markdown-as-Overlay Channel for Autonomous LLM Agents
N.E. Emilov (TU Delft - Electrical Engineering, Mathematics and Computer Science)
J.A. Pouwelse – Mentor (TU Delft - Electrical Engineering, Mathematics and Computer Science)
B. Nasrulin – Mentor (TU Delft - Electrical Engineering, Mathematics and Computer Science)
A.E. Zaidman – Graduation committee member (TU Delft - Electrical Engineering, Mathematics and Computer Science)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Communication protocols are hard to change: each node runs separately deployed code, so upgrading one means redeploying it on every node. Autonomous agents built on large language models must change the protocols they speak far faster than this allows. We ask whether such agents can implement and evolve a shared protocol from its description alone, without central infrastructure, and what channel design this requires.
We present DelftClaw, a decentralised channel on which each agent compiles a shared, self-verifying Markdown protocol into self-contained code and joins a community whose membership requires no custodian. Because a language model writes the code, this works only if independent compilations of a single description behave alike, which we test across three models and four protocols by running the compilations as communities against one another.
We find that this precondition holds: independent compilations reproduce a hand-written reference’s specified behaviour, and when agents evolve the protocols themselves, two compilations that share a model still converge on the same behaviour. Transmitting a protocol as a description can thus replace a network-wide redeployment with a single message.