The Big Picture
Most agent communication protocols already handle multi-turn conversations and structured data, but the field lacks agreed standards and essential privacy and governance features. Expect a move toward a layered, interoperable protocol stack rather than one single standard.
ON THIS PAGE
The Evidence
Agent-to-agent protocols (see Agent-to-Agent Protocol) almost always keep session state and mix free text with structured payloads, enabling multi-step plans and tool use. Schema negotiation—letting agents agree on data formats—is common and sometimes allowed to change at runtime. Discovery is usually centralized or static, with only one protocol using peer-to-peer discovery. Across the nine protocols studied, protections for privacy, compliance, and policy enforcement are largely missing.
Data Highlights
17 of 9 agent-to-agent protocols provide session state for multi-turn interactions.
27 of 9 protocols support hybrid payloads (text plus structured data); 7 of 9 permit multiple schema definitions and 2 of 9 allow schemas to evolve at runtime.
3Discovery methods: 4 of 9 use a centralized registry, 4 of 9 rely on static configuration, and only 1 of 9 (LMOS) uses decentralized peer-to-peer discovery.
What This Means
Engineers building systems of cooperating agents should use these findings to choose protocols that support multi-turn state and schema negotiation for robust coordination. Technical leaders and architects should plan for a layered, interoperable stack and add governance, privacy, and continuous evaluation features before production deployment. This can be guided by design patterns such as the Agent Service Mesh Pattern to ensure modular, policy-driven integration.
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1: Fundamental components of LLM agents.
Fig 2: Figure 2: Finalized Taxonomy: all dimensions and characteristics
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
The taxonomy is built at the application/protocol level and deliberately omits low-level implementation details, so practical engineering trade-offs (performance, cost) still need evaluation. The analysis covers nine protocols and may not capture newer proposals that appear after the study. Security and privacy gaps were identified but not exhaustively tested for every protocol—assume additional work is needed to meet compliance requirements. (Note related concerns in Zero-Click Data Exfiltration.)
Methodology & More
Researchers created a practical, hierarchical classification that organizes how language-based agents exchange information. Using an iterative, documented method for taxonomy design, they focused on protocols that connect an agent to another agent or to external services at the application level. The taxonomy was applied to nine concrete protocols to validate its coverage and to surface common patterns and gaps.\n\nKey findings show that agent-to-agent communication typically includes session state for multi-turn interaction and supports both free-text messages and structured data, so agents can both chat and call APIs. Schema negotiation—letting agents agree on data formats—is widely supported and sometimes allowed to change during runtime, which helps flexible collaboration. Most systems use centralized or static discovery, with only one peer-to-peer example. Importantly, few protocols include built-in privacy, compliance, or policy enforcement features, highlighting a major engineering and governance gap. The authors expect the ecosystem to move toward a federated, layered protocol stack—lightweight discovery, structured tool execution, and session-aware deliberation—rather than a single monolithic standard. (See Semantic Capability Matching Pattern and Planning Pattern for related design guidance.)
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Includes Alois Knoll, a recognized academic in robotics/AI—strong institutional reputation.