Overview

The Problem

The Agent2Agent (A2A) protocol governs how agents discover, authenticate, and exchange tasks with each other. The Model Context Protocol (MCP) governs how agents call tools and interpret results. Neither governs what happens at four critical boundaries:

1. The policy boundary. A2A’s AgentCard describes what an agent can do. It has no mechanism for what calling agents are permitted to do under the callee’s terms of service.

2. The outcome boundary. A2A’s task lifecycle ends at completed or failed. It has no vocabulary for whether the outcome served the principal, or whether the agents involved should behave differently next time.

3. The integrity boundary. A2A authenticates agents at the transport layer (TLS, Bearer tokens). It does not detect agents whose behavior has drifted after authentication, whether from adversarial compromise, RL misalignment, or model update side effects.

4. The principal boundary. Every protocol assumes the principal can see, hear, read, comprehend, decide in real time, and provide consent. For 1.3 billion people with disabilities and every adult whose capabilities are declining with age, this assumption is structurally exclusionary.

Four Protocols, Four Gaps

Protocol	Gap addressed	Phase
Sauvidya / PACE	Accessibility: can the principal participate?	Precondition
Anumati / ACAP	Consent: did the caller honor the callee’s terms?	Entry
Pratyahara / NERVE	Integrity: is the agent still trustworthy?	Ongoing
Phala	Welfare: did the action serve the principal?	Exit

Each protocol is independent. You can deploy any one without the others. They are also composable: deploying multiple protocols on the same AgentCard produces compound guarantees that no single protocol provides alone.

Design Principles

All four protocols share three architectural commitments:

1. Non-breaking A2A/MCP extensions. Every protocol uses the standard capabilities.extensions mechanism. No core spec changes are required. Agents that do not understand an extension ignore it.

2. Companion repositories with tests. Each protocol has a Python reference implementation with Pydantic types, runtime validators, and a pytest suite. What the paper specifies, the code enforces.

3. Honest validation framing. No protocol claims more than it has demonstrated. TLA+ model checking, numerical simulation, and runtime validators are clearly distinguished from empirical deployment evidence.

Author

Ravi Kiran Kadaboina — Independent Researcher, M.S. Computer Engineering, University of New Mexico, 2011.

All protocols are published as Zenodo preprints under Apache 2.0.