Agent-to-Agent (A2A) protocols let agents from different teams, companies, or platforms collaborate. The promise: an agent ecosystem where capabilities compose. The reality in 2026: early-stage standards, real interoperability for simple cases, custom integration for complex ones.
The capability discovery problem
Agent A needs to find agents that can do X. Solutions: registries (DNS-like), broadcast (capability advertisements), pre-shared knowledge. Each has trade-offs (centralization, scaling, security).
Request/response or streaming
Simple: request payload + response. Complex: streamed back-and-forth (clarification, partial results). A2A protocols often support both; pick by latency requirements and result size.
Authentication and trust
Workload identity (SPIFFE), signed messages, per-agent capability scopes. Agents are software, not users — service identity model fits better than user identity model. Pre-shared trust between organizations matters.
Result format negotiation
Calling agent prefers JSON; called agent emits markdown summaries. Negotiate at call time: 'I accept JSON, markdown'. Avoids brittle output parsing.
Where it's headed
Standards (Google A2A, Anthropic MCP-derived, others) are converging on JSON-RPC-like patterns over HTTP/2. Discovery is fragmented. Expect more consolidation in 2026-2027.