A2A (Agent2Agent protocol)

The missing inter-agent protocol in the wiki’s standards layer. The spoke already pages the context/skills standards — agentskills-spec (what an agent can do), agents-md (what it must know), and the cross-wiki MCP (how an agent reaches tools/data). A2A is the orthogonal axis: how agents talk to other agents. Originally built by Google, now donated to the Linux Foundation — the same open-standard-as-attractor pattern as agentskills-spec/agents-md.

What it standardizes

Agent-to-agent interoperability across vendors and frameworks — LangGraph, CrewAI, Semantic Kernel, ADK, and custom agents can collaborate without exposing internal memory, tools, or logic. Mechanisms:

Agent Cards — a published capability-discovery descriptor (what an agent can do / how to reach it). (Per-agent; ARD is the cross-org layer that indexes A2A agents and MCP servers so they can be found across domains in the first place.)
Tasks — the unit of work exchanged and delegated between agents; streaming + async for long-running work (the same durability concern, at the protocol layer).
Transport — JSON-RPC over HTTP.

A2A vs MCP — complementary, not competing

The cleanest framing this spoke has for its protocol layer:

MCP = agent → tool. How one agent accesses tools/APIs/data.
A2A = agent → agent. The “public internet” for agents to find and delegate to each other.
MCP runs inside an A2A ecosystem — different layers. So the orchestrator → subagent fan-out pattern, so far an in-process harness concern, gets a cross-vendor wire format: orchestration can now span agents from different builders.

Worked instance — cross-language agents (adk-a2a-cross-language)

The first concrete demonstration the spoke has: adk-a2a-cross-language builds a contract-compliance pipeline where a Python ADK orchestrator (SequentialAgent) drives a Go validator wrapped as a RemoteA2aAgent — the Go service is plain HTTP with no AI framework, A2A-compliant by protocol adherence alone. It exercises all three A2A mechanisms end to end (Agent Card at /.well-known/agent.json, JSON-RPC message/send, Task states), proving the cross-language, cross-framework claim with running code rather than spec prose — and adds a resilience note (agent unreachable → human MANUAL_REVIEW).

Why it matters here

Fills the inter-agent gap beside skills/context/tool standards — the spoke’s “standards layer” was missing the agent↔agent edge. With A2A, the standardization-reach open question sharpens: skills (agentskills-spec), context (agents-md), tools (MCP), and now interop (A2A) are four distinct open standards, each a candidate attractor — or a new front for fragmentation (ACP and vendor-native schemes still coexist).
Governance signal. Like agents-md (Linux Foundation) and agentskills-spec, a major vendor donating the spec to a neutral foundation is the recurring move to make a format an industry default rather than a house style.

The other branch — replace the substrate (pilot-protocol)

A2A’s design choice is to keep the human internet and add agent semantics above it: JSON-RPC over HTTP, agents reachable at ordinary URLs, discovery through a file at a well-known path. pilot-protocol takes the opposite branch — its own address space resolved by its own registry instead of DNS, a UDP session-layer protocol with NAT hole-punching so agents connect directly, and payments (USDC via x402) as a network primitive. Same layer of the problem (agent ↔ agent), opposite assumption about what to build on.

That makes the standards-layer question less “which format wins” and more “does the agent web run on the human one.” A2A’s bet is that HTTP’s infrastructure — DNS, TLS, proxies, every operator’s existing firewall rules — is worth more than the impedance mismatch costs. Pilot’s is that the mismatch is the point. The governance contrast is just as sharp: A2A is Google-originated and Linux-Foundation-governed, while Pilot is a venture-funded company operating the network it specifies.

Caveat

Spec/standard, not an outcome: adoption breadth is the open question (does A2A win the agent-interop layer or splinter against ACP / proprietary schemes?). Vendor-originated (Google), now foundation-governed.

model-context-protocol · agentskills-spec · agents-md · agentic-resource-discovery · pilot-protocol · agent-orchestration · adk · durable-agents · synthesis