Can agent protocols be efficient, versatile, and portable simultaneously?

The proliferation of agent communication protocols looks like noise until you classify it. This taxonomy, built through five iterations over nine actively-maintained open-source protocols, lands on five dimensions — counterparty, payload, interaction state, discovery mechanism, schema flexibility — and surfaces recurring architecture. All sampled agent-to-agent protocols pair hybrid payloads with session-state persistence; most support multiple predefined schemas, two negotiate schemas at runtime; decentralized discovery remains rare. The organizing principle is Marro's Agent Communication Trilemma: a protocol cannot simultaneously maximize versatility, efficiency, and portability.

That trilemma explains the design split cleanly. MCP-style context protocols enforce rigid schemas, stateless interactions, and structured payloads — buying portability and efficiency by eliminating token-heavy negotiation, ideal for predictable high-throughput agent-to-tool calls. Evolving-schema protocols like Agora and ANP buy versatility for open-ended multi-agent debate, at the cost of the negotiation overhead the rigid ones avoid. There is no free lunch; you pick a corner. The forward signal the authors read is convergence pressure toward protocols that unify agent-to-agent and agent-to-context communication — collapsing the tool-call channel and the peer-coordination channel into one substrate.

This frames the empirical findings elsewhere. Why do protocol-based tool integrations fail in production workflows? is exactly the efficiency-corner cost made concrete: MCP's portability comes with interpretation overhead that hurts determinism, a predictable consequence of the trilemma rather than a bug. And Can semantic capability vectors replace manual agent routing? is an attempt to solve the discovery dimension the taxonomy flags as underdeveloped (decentralized discovery is rare). The taxonomy's weakness is that classifying nine open-source protocols with demonstrable adoption may bias toward published, surviving designs; proprietary or short-lived protocols could occupy trilemma corners the sample never sees, and "convergence pressure" is an inference from a snapshot, not a demonstrated trajectory.