5. [Supplementary] Interoperability, Observability and Discoverability of Agentic Systems
5.1 Interoperability
Protocols like Agent-to-Agent (A2A) and Multi-Protocol Communication (MCP) are central to ensuring the interoperability of an Agentic AI framework. These protocols are designed to facilitate the interaction, exchange of data, and alignment of goals between different autonomous agents or systems.
5.1.1 A2A
The Agent-to-Agent (A2A) protocol shall govern the information exchange, collaboration, and negotiation among individual agents. In an Agentic AI system, agents are typically tasked with independent actions but must be able to coordinate when necessary. This means that agents need to understand each other's intentions, capabilities, and status.
Key aspects of A2A include: - Information Exchange: Ensures that agents can send and receive relevant information (e.g., sensor data, action requests, state updates) to other agents in the system. - Collaboration and Negotiation: Enables agents to coordinate on joint tasks. This is particularly important in domains like multi-robot systems or in financial markets where agents need to negotiate, compromise, and adapt based on the actions and needs of other agents. - Conflict Resolution: A2A communication can also provide mechanisms for agents to resolve conflicts when their individual goals are misaligned such as via hierarchical negotiation or probabilistic consensus.
5.1.2 MCP
MCP focuses on standardizing how agents communicate with tools that enable additional functionality. Given that AI systems may involve a diverse range of devices, software components, and networks, agents need the ability to communicate across these varied contexts while being abstracted from the actual implementation of the endpoints they are communicating with. MCP plays a vital role in this by ensuring that the communication between different agents and systems, regardless of the protocol or infrastructure they use, remains coherent and standardized. Key aspects of MCP include: - Protocol Bridging: MCP allows agents to interface with multiple communication protocols and standards (e.g., HTTP, MQTT, WebSockets, etc.) without requiring a single protocol to dominate. This allows agents to interact with legacy systems or third-party services that might use different communication methods. - Flexibility and Compatibility: MCP ensures that the agentic framework remains adaptable to new protocols or systems as they emerge. This means that the system can quickly integrate new technologies or standards without major overhauls. - Data Translation and Contextualization: MCP can include mechanisms for translating and contextualizing data between agents that may interpret or process information differently. For example, one agent might deal with raw sensor data while another operates on higher-level abstractions, and MCP ensures that the data is appropriately transformed between these levels.
5.2 Observability
Observability is paramount in agentic AI systems due to their autonomous nature and the complexity of their operations. End-to-end visibility into agent execution (how agents reason, where they access information, and what actions they take) provides data essential for transparency, optimization, and detecting security anomalies.
5.2.2 Key Observability Pillars
Following the principles for A2A, Agents shall - Participate in Distributed Tracing via OpenTelemetry. - Provide comprehensive logging at both the client and server side. - Provide operational metrics via Prometheus.
The requirements establish the three foundational pillars of observability: Tracing, Logging, and Metrics. Building on these pillars, Appendix 7.5 provides a comprehensive observability requirements framework that translate them into actionable guidance across six operational categories: agent-level observability, agent performance, user interaction/behavior, system performance/health/reliability, safety and security compliance, and product lifecycle/governance. For each category, the framework identifies specific risks, assesses their potential impacts, recommends controls, and details the metrics/logs/traces needed. Please refer to Appendix 7.5 for more details on each pillar and the observability requirements framework.
5.3 Discoverability
An Agent Registry serves as a centralised directory for all AI agents within a system or organisation. Its primary purpose is to enable seamless Agent Discovery, allowing various components and users to efficiently locate, understand, and interact with available agents. A2A provides a framework to implement this.
Please see Appendix 7.6 for more details.