A2A (Agent-to-Agent) protocol matters primarily when agents cross ownership or trust boundaries — different teams, vendors, or deployments. Within a single trust boundary, orchestration frameworks plus MCP are usually sufficient. The post explains A2A's core primitives (Agent Cards, Tasks, Messages, Artifacts), its security model (OAuth 2.0, mTLS, Signed Agent Cards), and the unresolved delegation gap where A2A defers to external identity providers. MCP and A2A are complementary: MCP connects agents to tools and data, while A2A connects independently deployed agents to each other. Production adoption of A2A remains hard to verify, partly because many 'multi-agent' systems are really single orchestrators calling tools sequentially. The practical guidance: establish tool access with MCP first, add A2A only when a genuine deployment or ownership boundary exists. Redis is promoted as the underlying context/memory layer for agent state regardless of which protocol is used.
