Agent-to-Agent Protocol (A2A) and the Java SDK

An airhacks.fm conversation with Kabir Khan (@kabirkhan) about: Discussion about the A2A (Agent-to-Agent) protocol initiated by Google and donated to the Linux Foundation, the A2A Java SDK reference implementation using quarkus, the Java SDK development accepted by Google, comparison of python and Java expressiveness and coding practices, the concept of an agent as a stateful process versus a tool as a stateless function call, the agent card as a JSON document advertising capabilities including supported protocols and descriptions and input/output modes and examples, the three wire protocols supported: JSON RPC and HTTP+JSON (REST) and grpc, the proto file becoming the single source of truth for the upcoming 1.0 spec, the facade/adapter pattern for the unified client across protocols, the agent executor interface with request context and event queue parameters, the distinction between simple message interactions and long-running multi-turn tasks, tasks as Java Records containing conversation history with messages and artifacts, message parts including text parts and data parts and file parts, task lifecycle with task IDs and context IDs for stateful conversations, the event queue as internal plumbing for propagating task updates, the separation between spec package (wire protocol entities) and server package (implementation details), the task store as a CRUD interface with in-memory default and database-backed implementations in extras, replicated queue manager using microprofile reactive messaging with Kafka for kubernetes environments, building A2A agents without any LLM involvement for simple use cases like backup systems, the role of LLMs in creating prompts from task messages and context, the agentic loop and the challenge of deciding when an agent's work is complete, the relationship between MCP (Model Context Protocol) for tool access and A2A for agent-to-agent communication, the possibility of wrapping agent calls as MCP tools, memory management considerations with short-term and long-term memory and prompt size affecting LLM quality, the distinction between the bare reference implementation and Quarkus-specific enhancements like annotations and dependency injection, upcoming 1.0 release with standardized Java records for all API classes and improved JavaDoc, protocol extensions including the agent payment protocol and GUI snippet extensions using template engines, authentication support with OAuth2 tokens and API keys and bearer tokens, the authenticated agent card containing more information than the public agent card, authorization hooks being discussed for task-level access control, the API and SPI segregation suggestion for better clarity between spec and implementation Kabir Khan on twitter: @kabirkhan