Google A2A Protocol Explained: Multi-Agent in 2026

Google A2A Protocol Explained: The 2026 Guide to Agent-to-Agent Communication

At a Glance (Updated May 2026): Google's Agent-to-Agent (A2A) protocol is an open standard that lets AI agents discover, authenticate, and communicate with other agents — across vendors, platforms, and clouds. According to TURION.AI's April 2026 launches recap, Google pushed A2A to v1.0 at Cloud Next '26 (April 2026), making multi-agent orchestration a first-class primitive across Vertex AI. A2A is the agent-to-agent counterpart to MCP's agent-to-tool standard — together they form the wire-level infrastructure for production multi-agent systems. This guide explains what A2A is, how it works, and how Fleece AI's hierarchy aligns with it.

What Is the A2A Protocol?
Why A2A Exists
How A2A Works
A2A vs MCP: How They Complement Each Other
What Changed at v1.0 (Cloud Next '26)
How Fleece AI's Hierarchy Aligns With A2A
7 Multi-Agent Workflows A2A Unlocks
Limitations and Open Questions
FAQ

Key Takeaways

The Agent-to-Agent (A2A) protocol is an open standard that defines how independent AI agents discover each other, authenticate, exchange messages, and delegate sub-tasks — regardless of underlying model or vendor.
Google announced A2A in 2024 and pushed it to v1.0 at Cloud Next '26 in April 2026, with first-class support across Vertex AI, according to coverage of the launch.
A2A complements MCP: MCP standardizes agent-to-tool calls; A2A standardizes agent-to-agent collaboration. Production multi-agent systems will use both.
Core A2A primitives are the Agent Card (a discovery document — capabilities, auth, endpoints), Tasks (units of work delegated between agents), and Messages (the conversational layer for streaming updates and clarifications).
Fleece AI's hierarchical agent system implements an A2A-style design — agents discover sub-agents, delegate tasks, and receive results — over a managed runtime so teams get the architecture without writing protocol code.

What Is the A2A Protocol?

The Agent-to-Agent (A2A) protocol is an open specification, originally proposed by Google in 2024 and shipped to v1.0 at Cloud Next '26 in April 2026, that defines how AI agents collaborate with other agents. It is the protocol equivalent of "USB for agent teams" — agents from different vendors, running on different models, hosted on different clouds, can cooperate over a single standardized contract.

A2A solves a problem that emerged the moment teams started building multi-agent systems: every framework had its own internal messaging format. LangGraph agents talked to LangGraph agents. AutoGen agents talked to AutoGen agents. Vertex AI agents talked to Vertex AI agents. Cross-framework? Cross-vendor? Cross-cloud? You wrote glue.

A2A removes the glue.

Why A2A Exists

The 2026 reality is that no single vendor will house all your agents. According to Gartner's 2026 Hype Cycle for Agentic AI, governance and interoperability are the deciding deployment criteria — and you can't have either if your CRM agent (vendor A) cannot talk to your finance agent (vendor B). A2A is the protocol layer that makes vendor-mixed deployments tractable.

There's also an enterprise-architecture argument. Different teams in a company will buy or build agents for their domains: a sales team standardizes on Salesforce's agentic stack; a finance team picks SAP's; engineering builds custom. Without A2A, those agents are islands. With A2A, the sales agent can ask the finance agent "is this customer in good standing?" without anyone in IT writing custom integration code.

How A2A Works

A2A defines a JSON-over-HTTPS protocol with three primitives:

Agent Card

A discovery document published by an agent (typically at /.well-known/agent.json) describing:

The agent's name and description.
Its capabilities (skills it advertises).
Authentication requirements (OAuth 2.1, API key, etc.).
Streaming and async support.
Endpoint URLs for tasks and messages.

A client agent reads the card to know what the server agent can do.

Tasks

A unit of delegated work. The client agent sends a task to the server agent (a POST /tasks request); the server returns a task ID, processes the work, and the client polls or subscribes for updates. Tasks have lifecycle states: submitted → working → input-required → completed (or failed / canceled).

Messages

The conversational layer. While a task is running, the two agents can exchange messages — clarifications, partial outputs, confirmations. This is what makes A2A different from a one-shot RPC: agents have a conversation about the task as it progresses.

Auth is OAuth 2.1 by default, with bearer tokens or mTLS as options. Streaming uses Server-Sent Events.

A2A vs MCP: How They Complement Each Other

Dimension	MCP	A2A
Best For	Agent-to-tool integration	Agent-to-agent collaboration
Created by	Anthropic (Nov 2024)	Google (2024, v1.0 April 2026)
Primitives	Tools, Resources, Prompts	Agent Card, Tasks, Messages
Use case	"Fetch this row from the DB"	"Delegate research to the researcher agent"
Conversation model	Request/response	Multi-turn, async
Discovery	Server lists capabilities	Agent Card
Streaming	Yes	Yes (SSE)
Auth standard	OAuth 2.1	OAuth 2.1
Vendor support	All major	Google + ecosystem (growing)
Pricing	Free/open	Free/open

The right framing: an A2A-compatible agent uses MCP to call its tools and uses A2A to call other agents. They're complementary infrastructure layers, not competitors.

Build A2A-style hierarchies on Fleece AI — manager agents delegate to specialists out of the box. Start at fleeceai.app.

What Changed at v1.0 (Cloud Next '26)

According to coverage of Cloud Next '26, the v1.0 milestone shipped with three production-readiness improvements:

First-class Vertex AI integration. Any agent built on Vertex AI publishes an Agent Card automatically and can be discovered by any other A2A-compatible client without manual wiring.
Hardened auth. OAuth 2.1 with PKCE is now the default, replacing the looser bearer-only scheme of the early drafts.
Streaming task updates. SSE-based streaming for partial outputs is part of the spec rather than an extension.

The v1.0 stamp matters because it signals API stability — enterprises can now build on A2A without expecting breaking changes every quarter.

How Fleece AI's Hierarchy Aligns With A2A

Fleece AI's hierarchy system implements an A2A-style architecture without requiring you to think about wire formats:

Agent Card → agent profile. Every Fleece AI agent has a profile (name, description, capabilities, custom skills, knowledge files). A manager agent reads its team's profiles via the list-sub-agents tool.
Tasks → delegation tools. A manager agent delegates a sub-task to a child agent (verified by parent ownership). The child agent runs and reports back.
Messages → audit log. Every delegation, report, prompt update, and broadcast is logged to an inter-agent messages table you can inspect.

The runtime enforces the operational primitives that the protocol leaves to implementations: rate limits (5 prompt updates per agent per day), depth bounds (max 3 levels), cycle detection, and cross-user authorization checks. The result is "A2A-compatible architecture, managed runtime."

7 Multi-Agent Workflows A2A Unlocks

1. Cross-Vendor Research Squad

A finance agent (vendor A) asks a research agent (vendor B) to gather competitor pricing — across A2A. No custom integration. See the multi-agent guide.

2. Hierarchical Outbound Sales

A CEO agent delegates strategic asks to a Sales Director agent, who delegates to regional SDR agents — exactly the hierarchy pattern.

3. Cross-Cloud Incident Response

A monitoring agent on AWS detects an incident and delegates investigation to an analytics agent on GCP. Each agent stays in its native cloud.

4. Departmental Manager Agents

HR's onboarding agent delegates "create laptop ticket" to IT's helpdesk agent. Each manager agent stays in its department's stack.

5. Tenant-Isolated Agent Teams

In multi-tenant platforms, each tenant's manager agent talks to that tenant's specialists — A2A's auth model maps cleanly to tenant boundaries.

6. Document Processing Pipeline

Extractor → classifier → router → specialist (legal vs finance vs HR). Five agents, possibly from different vendors, coordinate via A2A.

7. Customer-Facing Agent Federation

Your customer-facing agent talks to your partner's customer-facing agent to fulfill a cross-company workflow (ship + invoice + notify). Without A2A, this is a multi-quarter integration project.

A2A also pairs naturally with long-context AI agents — when a manager agent holds millions of tokens of working memory inline, delegated sub-tasks across A2A return into a context that can hold the full delegation history rather than a lossy summary.

Limitations and Open Questions

Vendor coverage is uneven. v1.0 shipped with Vertex AI as the reference implementation. OpenAI and Anthropic A2A surfaces are evolving; expect parity over the next 6–12 months.
Discovery still benefits from registries. /.well-known/agent.json works for known endpoints; finding the right agent for an unknown task still requires a directory layer.
Cost accountability across agents is hard. When agent A delegates to agent B and B does heavy work, who pays? Federated billing is an open problem.
Trust and policy. Agent A trusting agent B's output — especially across organizations — needs richer policy than auth tokens. This is on the 2026/2027 roadmap.

FAQ

How does A2A differ from a webhook?

A webhook is a one-way HTTP push for events. A2A is a bidirectional, multi-turn, stateful conversation between agents — designed for tasks that take seconds to hours and may require clarifications mid-task. Webhooks are notifications; A2A is collaboration.

Do I need A2A if I'm using a single platform like Fleece AI?

Not for your internal multi-agent workflows — Fleece AI's hierarchy system handles that natively. A2A becomes valuable when your agents need to talk to agents outside your platform (a partner's agent, a vendor's agent, a different cloud's agent).

Is A2A only for Google Cloud?

No. A2A is an open spec. Google leads the implementation and ships it natively on Vertex AI, but the protocol is publicly specified and any vendor can support it. Adoption beyond Google is growing through 2026.

How does A2A handle authentication between two organizations?

OAuth 2.1 with PKCE is the v1.0 default. The calling agent obtains a token scoped to the receiving agent's capabilities; the receiving agent validates and enforces fine-grained scopes. mTLS is supported for higher-trust deployments.

Can A2A and MCP both run on the same agent?

Yes — and they should. An A2A-compatible agent uses MCP to call tools and A2A to call other agents. Each protocol covers a different layer of the stack.

The Bottom Line

A2A v1.0 is to multi-agent systems what HTTP was to the early web — a boring, foundational standard that suddenly makes a thousand previously-impossible workflows trivial. Combined with MCP for tools, it forms the production wire-format for multi-agent AI in 2026. If you're building agent teams that span vendors, clouds, or organizations, A2A is the layer you'll be writing on top of for the rest of the decade.

Model Context Protocol Explained — the agent-to-tool standard
Multi-Agent AI Systems Guide — the use case
AI Agent Hierarchy & Delegation Guide — practical hierarchy
What Is an AI Agent? — pillar definitions
What Is Fleece AI? — overview
Best Autonomous AI Agents 2026 — platform roundup
Hermes Agent vs Fleece AI — open-source vs managed
OpenAI Workspace Agents vs Fleece AI — head-to-head

Build A2A-style hierarchies on Fleece AI — managers, specialists, auto-delegation, audit logs, all managed.