RAMP: A Protocol for Human Oversight of Autonomous Agents

Autonomous AI agents are rapidly transitioning from research prototypes to production systems that take consequential actions on behalf of human principals: spending money, sending communications, and modifying infrastructure. The agent protocol ecosystem has standardized tool access (Model Context Protocol, MCP) and inter-agent coordination (Agent-to-Agent Protocol, A2A). However, the Agent-to-Human supervisory channel lacks an open protocol integrating mandatory temporal safety bounds, declarative multi-tenant governance, and cryptographic auditability. We present RAMP (Remote Agent Monitoring Protocol), a transport-agnostic, cryptographically verifiable protocol that fills this gap. RAMP defines five primitives (telemetry, notifications, action requests, governance policies, and audit trail) organized into three functions: Observe (what is the agent doing?), Decide (should the agent do this?), and Govern (what is the agent allowed to do?). The protocol specifies a state-machine agent lifecycle, HMAC-SHA256 message integrity with replay protection, a declarative governance engine with six normative rule types and formal precedence ordering, and a tamper-evident hash-chained audit trail. We provide a formal specification, a reference gateway and Python SDK achieving Level 3 conformance (full governance and audit), and a positioning analysis demonstrating RAMP as complementary to MCP and A2A. The specification, implementation, and SDK are open-source under the Apache 2.0 license.