Symplex Protocol – semantic intent vectors for AI agent communication (Go, v0.1)

  I built Symplex because I kept running into the same wall: MCP and similar
  protocols require agents to agree on a schema before they can communicate.
  That works fine for controlled environments, but breaks down when you want
  agents from different teams or systems to discover and collaborate dynamically.

  Symplex replaces JSON tool-call schemas with float32 intent vectors — semantic
  embeddings that represent what an agent *wants to do*. Capability matching uses
  cosine similarity, so an agent advertising "summarize document" can match a
  request for "condense text" without any shared vocabulary or pre-agreed schema.
  Transport is libp2p, identity is Ed25519 DIDs, and negotiation is spontaneous
  rather than orchestrated.

  The architecture is a P2P mesh rather than hub-and-spoke. Agents publish
  capability vectors, subscribe to intent streams, and negotiate directly.
  Routing decisions are made semantically, not by name-matching against a
  static registry.

  Honest caveat: this is v0.1, MIT-licensed, written in Go. Tests pass and the
  core protocol works, but it's a proof of concept — not production-ready. The
  embedding model is pluggable but you need to wire one in. I'm releasing now
  because I want feedback on the protocol design before building further.

  Repo: https://github.com/olserra/symplex