Agent Vector Protocol (AVP) Specification
What it is
AVP is an open specification for a binary protocol that enables LLM agents to communicate via latent representations (KV-cache and hidden states) instead of serialized text. It defines a handshake mechanism to determine compatibility, allowing for direct latent transfer between same-model agents, vocabulary-mediated projection for cross-model agents, or a fallback to JSON text.
The problem it solves
Standard agent-to-agent communication relies on text serialization, which discards computed KV-cache and hidden states, forcing redundant and slow autoregressive generation for every message exchange.
What you could use it for
- Accelerating multi-agent pipelines by eliminating redundant autoregressive decoding between agents using the same model
- Enabling communication between agents running different LLM architectures via zero-training vocabulary projection
- Integrating latent state transfer into existing agent frameworks like LangChain, CrewAI, and AutoGen
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.