The latent communication protocol. A network protocol that lets LLM agents served by distinct inference engines exchange latent (hidden-state) messages over real links — designed as research: a spec, a reference implementation, and measurements on real links.
This is the third chapter of a series on latent inter-agent communication:
- Whitepaper 1 — training-free quantization of the latent channel
(
github.com/Neetx/latent-channel-compression). - Whitepaper 2 — serving latent multi-agent systems in an inference
engine (
github.com/Neetx/latent-link-vllm): vLLM fork + library, bit-parity 250/250, real bytes at ×8 with no degradation, ~1400 problems/hour on one RTX 5070 Ti. - Whitepaper 3 (this repo) — lift the channel onto the network. The MAS-ness moves between the engines, not inside; each engine goes back to serving a single model.
AGENTS.md— rules, environment, toolchain (binding for agents).RESEARCH.md— established results (inherited anchors + wp3 findings).ROADMAP.md— status and work queue.REPRODUCING.md— re-run every experiment (reports 01–16).DEVELOPMENT.md— the reference implementation (package → spec map).docs/design/— the spec (protocol_spec_v0.md), the plans, decisions.docs/reports/— numbered verification reports +README.md(the family map).writeup/— the whitepaper-3 paper (main.tex; figures generated from the committed artifacts).
All three experiment families complete; paper drafted. 16 numbered reports; 70 tests.
- Protocol validation (done — reports 01–10) — the reference
implementation (
latent_protocol: frame, shm/TCP/QUIC transports, manifest, handshake, codecs, conformance, orchestrator) validated bit-exact and end-to-end with the real models via an HF-backend runtime; canonical Math-500 accuracy at n=250 (76.0%); the transport/RTT/codec systems characterization. - Distinct vLLM engines over the protocol (done — reports 12–14) — the thin fork (2 primitives) + orchestrator driving three separate vLLM engines: canonical-level accuracy at n=50 (86.0%, statistically compatible with the HF reference on the same subset), ×4.9 the HF-floor throughput, and the RTT law on real engines (compute-bound LAN, RTT-bound WAN).
- Channel & payload generality (done — reports 15–16) — the text-channel baseline (text is 58× smaller on the wire; the latent channel's payoff is trained collaboration, not bytes) and a second binding (a KV-concat channel carried bit-exact with no wire-layer changes).
Whitepaper 4 (production-grade engine) is a separate repo and session, like wp1/wp2 — out of scope here.
Apache-2.0 (consistent with whitepaper 2 and vLLM). See LICENSE.