Three live agent protocols, three different jobs. Here is the comparison and which one you actually need this quarter.

There are three agent protocols you need to know in mid-2026: MCP, A2A, and ANP. They are not competing. They solve different problems, and most teams we work with end up using at least two of them.
The confusion comes from vendors who use "agent protocol" as if it meant one thing. It does not. Picking the wrong one for the job is how you end up with a six-month migration.
If MCP is USB-C, A2A is email between companies, and ANP is the DNS plus handshake that lets you find the right agent in the first place.
MCP is Anthropic's protocol, open-sourced in November 2024. It standardizes how a single agent reads files, calls APIs, queries databases, and runs prompts. We covered it in detail in the prior post. The key facts: 73% of production agent teams use it as of Q2 2026, every major model supports it, and 4,000+ public MCP servers exist.
MCP is the layer closest to the model. When your agent needs to "look something up" or "do something," MCP is the protocol that makes that happen.
A2A is Google's contribution, announced at Cloud Next in April 2025 and now at v1.0 with 50+ enterprise partners including Salesforce, SAP, ServiceNow, MongoDB, and Workday. The protocol defines how one agent hands a task to another agent owned by a different team, company, or vendor.
The use case A2A targets is real: your procurement agent needs to talk to your supplier's inventory agent. Both are autonomous. Neither has a human in the loop. Neither exposes a stable REST API. The agents need to discover each other's capabilities, negotiate an interface, exchange structured tasks, and stream results back.
A2A defines:
Critically, A2A is agent-opaque. Two agents do not need to share model providers, frameworks, or implementation details. They just need to speak A2A.
The big enterprise vendors adopted A2A because their customers demanded cross-vendor agent interoperability. Salesforce Agentforce talks to ServiceNow agents. SAP agents hand off to Workday agents. None of them had a clean way to do that before.
ANP came out of an open working group in 2025 and is now stewarded by the Linux Foundation's AI and Data group. It is the youngest of the three and the most ambitious. Where A2A assumes you already know which agent to talk to, ANP is about discovery and negotiation on open networks.
ANP defines:
The pitch: in 2027 or 2028, there will be millions of agents on the open internet doing specialized work. You will not hardcode which agent to call. Your agent will discover them by capability, verify their identity, negotiate a contract, and execute.
In 2026, ANP is early. There are fewer than 200 production ANP deployments we know of, mostly in financial services and supply chain. The spec is still moving at v0.6. But it is the protocol to watch for any cross-organization agent orchestration that cannot rely on a vendor lock-in.
| Dimension | MCP | A2A | ANP |
|---|---|---|---|
| Solves | Agent to tools | Agent to agent (known) | Agent discovery on open networks |
| Spec owner | Anthropic (open) | Google + Linux Foundation | Linux Foundation AI and Data |
| Spec version (Q2 2026) | 2025-11-25 (stable) | 1.0 (stable) | 0.6 (draft) |
| Primary use case | Internal tool integration | Cross-vendor agent handoff | Open agent marketplaces |
| Latency overhead | 40-80ms per tool call | 100-300ms per task handoff | 500ms-2s for discovery plus handoff |
| Production deployments | 4,000+ public servers | 50+ enterprise partners, ~2,000 active | ~200 deployments |
| Backed by | Anthropic, OpenAI, Microsoft | Google, Salesforce, SAP, ServiceNow | Linux Foundation, open working group |
| Best for | Any team building production agents | Multi-vendor agent ecosystems | Discovery-heavy cross-org workflows |
| Mature enough for production in 2026 | Yes | Yes (with caveats) | Pilot only |
For a 50 to 200-person ops team, the answer is almost always: start with MCP, add A2A only if you need cross-vendor handoff, ignore ANP until 2027.
Here is the decision rule we use with clients.
MCP only (90% of teams): If your agent works inside your company, talking to your tools, you need MCP. You do not need A2A. You do not need ANP. Stop reading and deploy MCP.
MCP plus A2A (8% of teams): If your agent has to hand off tasks to an agent you do not control, you need A2A. The most common case in 2026 is procurement and supplier coordination. The second most common is HR, where Workday agents and ADP agents need to coordinate on payroll changes.
MCP plus A2A plus ANP (2% of teams): If you are building an open agent marketplace or you are in financial services / supply chain with dozens of counterparties, you need all three. ANP for discovery, A2A for handoff, MCP for the tool layer inside each agent.
One of our clients, a 120-person UK logistics company, runs a freight booking agent. Here is the actual stack.
Total setup took 18 days, including two weeks of MCP server hardening and a week of A2A integration testing. The agent handles 200+ bookings per day with 92% straight-through processing, per the client's Q2 ops report.
The protocols are not a religious choice. They are a stack.
The mistake we keep seeing is teams picking one protocol and trying to use it for everything. MCP does not solve agent-to-agent handoff. A2A does not solve tool integration. ANP does not solve either yet, it solves discovery. Pick the right tool for the right job.
We are watching the agent protocol stack settle into the same shape as the early web's protocol stack. MCP is HTTP for the agent world. A2A is SMTP. ANP is DNS. Each has a clear job, each emerged from a different community, and they compose.
The companies that win on AI automation in 2026 and 2027 are the ones that treat these as infrastructure, not as products. You do not argue about HTTP versus SMTP. You use both. Same here.
Book a discovery call when you are ready to scope one high-impact workflow for production delivery.
Spread the word on your network or copy the link.