In 2024 we built single AI agents. In 2025 we built single agents with tools. In 2026 we are building systems of agents that delegate work to each other. The protocol that makes this possible is A2A (Agent-to-Agent), launched by Google in April 2025 and now supported by every major AI platform.
If MCP is how an AI talks to your tools, A2A is how AI agents talk to each other.
The Problem A2A Solves
Imagine an orchestrator agent that handles customer support. A user asks: "Can you tell me where my order is and refund the shipping fee?" The orchestrator has two specialist agents available: one that owns logistics data and one that handles refunds. In 2024, you would hand-code the routing. In 2026, the orchestrator can discover, delegate to, and coordinate with both specialist agents using A2A.
Three things had to be standardized for this to work:
- Discovery. How does the orchestrator find specialist agents and learn what they do?
- Capability negotiation. How do agents agree on what one is asking the other to do?
- Stateful conversation. A handoff is not a single request. It is a multi-turn dialogue with shared context.
A2A defines a JSON-based message format and HTTP transport that handles all three.
How A2A Compares to MCP
Easy way to remember it:
- MCP = AI talks to tools (deterministic functions, databases, APIs).
- A2A = AI talks to other AIs (each side is itself reasoning, negotiating, sometimes calling its own MCP tools).
In a typical 2026 production system, both protocols run side by side. The orchestrator agent uses A2A to delegate to specialists, and each specialist uses MCP to call its own internal tools.
A Real Example: Sales Pipeline Automation
One of our clients runs a B2B sales operation with three internal AI agents:
- Lead Qualifier reads inbound emails, scores leads, decides what stage they belong in.
- Researcher takes a qualified lead and gathers public info (LinkedIn, company size, recent news).
- Drafter writes a personalized first reply tailored to the lead's role and company.
Before A2A, this was a hardcoded chain. Now the orchestrator decides which agents to invoke based on the inbound email. If a lead is already enriched, it skips the Researcher. If the lead is too cold, it skips the Drafter. The chain is dynamic and explainable.
Anatomy of an A2A Call
An A2A request looks like this (simplified):
POST /agents/researcher/messages
Authorization: Bearer <agent-jwt>
{
"task_id": "task-7842",
"from_agent": "orchestrator",
"intent": "enrich_lead",
"input": {
"name": "Adaeze Okafor",
"email": "[email protected]",
"company": "Lagos Logistics Ltd"
},
"context": {
"conversation_id": "conv-9923",
"priority": "normal"
}
}The Researcher returns a structured response with the enrichment data plus metadata about confidence, sources, and any clarifying questions. If the Researcher is unsure, it can ask a follow-up via the same conversation_id.
Where A2A Breaks Down
Things to know before you commit to a multi-agent design:
- Cost compounds quickly. Two agents talking is two LLM calls per round trip. Five agents talking can be 20+ calls. Track token spend per workflow before scaling.
- Failure modes get weird. When agents disagree or loop, you need explicit timeouts and circuit breakers. We always add a hard "max 3 rounds of negotiation" cap.
- Observability is critical. Add request IDs to every message and log the full conversation graph. When something goes wrong at 2am, you will need it.
- Not every problem needs multiple agents. A single agent with the right tools is simpler, faster, cheaper, and easier to debug. Use multi-agent only when capabilities clearly separate.
When to Use Multi-Agent (and When Not to)
Use multi-agent A2A architectures when:
- Specialist agents need different tools, permissions, or model sizes.
- The work has clear stages with handoffs (qualify, research, draft, review).
- You want different teams or vendors to own different agents.
Stick with a single agent when:
- The task fits in one prompt and one tool set.
- Latency matters more than modularity.
- You are just starting and have not yet seen which capabilities will diverge.
Where to Go From Here
If you are designing an AI workflow that has felt too complex for one agent, A2A is the modern answer. Our AI automation practice has shipped multi-agent systems for sales, support, and operations clients. We will scope your workflow, identify the right specialist boundaries, and build the orchestration layer in production. Tell us your use case and we will respond within 24 hours.