Agent Transport Protocol: Asynchronous Store-and-Forward Messaging for Autonomous AI Agents

This document specifies the Agent Transport Protocol (ATP), an asynchronous store-and-forward messaging protocol for autonomous AI agents. ATP enables agents to transmit themselves -- including state, context, capabilities, and cryptographic identity -- between agent runtimes across network boundaries. The protocol draws on the operational model of the Simple Mail Transfer Protocol (SMTP) [RFC5321] but is purpose-built for agent-to-agent communication where the agent itself is the payload. ATP provides: (1) asynchronous delivery with store-and-forward semantics, (2) cryptographic identity verification at each relay hop, (3) trust scoring and policy enforcement at ingress, (4) capability negotiation between sending and receiving runtimes, and (5) tamper-evident envelopes with end-to-end integrity protection. The protocol is transport-agnostic and operates over TCP, TLS, QUIC, or any reliable ordered stream. ATP is designed to interoperate with existing agent frameworks including Google A2A, the Model Context Protocol (MCP), and FIPA ACL, while addressing the fundamental limitation of synchronous RPC-based agent communication: the requirement that both endpoints be simultaneously available.