Which organization governs MCP, A2A, and ACP as of 2026?
MCP, A2A, and ACP each have separate governance in 2026: MCP under Anthropic, A2A under the Linux Foundation, ACP under IBM Research / BeeAI. No single umbrella body; interop comes from working groups.
Think of three companies that each invented their own kind of electrical plug. At first, each company controlled its own design. That worked fine alone, but devices could not mix and match. Eventually they all agreed to hand their designs to a neutral standards body so the plugs would play nicely together. The agent protocols followed the same arc. MCP came from one company, the agent to agent protocol from another, and a third interop effort from elsewhere. By 2026 they sit under a shared, vendor-neutral home so any tool, agent, or host can interoperate without picking a single vendor's side.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
This question looks like trivia, but it tests whether you can hold the three agent protocols apart and describe who actually stewards each one in 2026.
Three protocols anchor the 2026 agent landscape. MCP, the Model Context Protocol, connects an agent to tools, data, and prompts. The agent to agent protocol lets separate agents discover and talk to one another. ACP, the Agent Communication Protocol, is a third effort aimed at richer agent messaging. Each began under a different vendor or community, and the governance picture has not collapsed into a single body.
The expected answer is that each protocol is governed independently: MCP under Anthropic's open spec, A2A donated to the Linux Foundation in 2025, and ACP under IBM Research and its BeeAI community. This deep dive explains where each protocol actually lives today, why only A2A has made the canonical vendor to foundation move so far, and what the three protocols actually do so you never confuse their layers.
The three protocols and why they are not rivals
The single biggest mistake is treating MCP, A2A, and ACP as three contenders for one slot. They are not. They sit at different layers of agent infrastructure and frequently run together. An interviewer who hears you describe them as competing for the same role learns immediately that you have not built with them.
MCP is the agent to tools layer. A host application running an LLM uses MCP to discover and invoke tools, read resources, and pull prompt templates from external servers over JSON-RPC. It answers a single question: how does one agent reach the outside world? The boundary it standardizes is the one between a single agent and its capabilities.
The agent to agent protocol, originated by Google, is the agent to agent layer. It lets distinct agents, possibly built by different teams or vendors, find each other, advertise what they can do, and exchange tasks or messages. The boundary here is between peers, not between an agent and a tool. ACP is a third effort aimed at broad agent interoperability, often described as REST-oriented and coming from a separate community lineage, sometimes associated with IBM and the BeeAI ecosystem.
Because they target different boundaries, a real system can use all three at once. An agent might use MCP to call tools, the agent to agent protocol to delegate a subtask to a peer agent, and an interop layer to bridge ecosystems. The complementary framing is the senior signal, and it is the lens that makes the governance question easy to reason about.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Protocol | Problem it solves | Original vendor |
|---|---|---|
| MCP | Connect one agent to tools, resources, and prompts | Anthropic (open spec, modelcontextprotocol working group) |
| A2A | Let separate agents discover and message each other | Google -> donated to Linux Foundation (2025) as Agent2Agent |
| ACP | Broader agent interoperability, REST-oriented | IBM Research / BeeAI project |
| Governance (2026) | Three independent tracks; no single umbrella body | Anthropic + Linux Foundation + IBM/community |
Real products, models, and research that use this idea.
- Anthropic continues to publish MCP spec revisions through the modelcontextprotocol working group on GitHub, with no donation to a neutral foundation as of 2026.
- Google donated A2A to the Linux Foundation in 2025 as the Agent2Agent project, following the same vendor to neutral foundation pattern as Kubernetes under the CNCF.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy does vendor-seeded then neutrally governed beat a standard born inside a committee?
Velocity first: one owner ships a working spec fast and proves demand. Neutrality comes later, once adoption needs trust that no rival can gate the standard.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Assuming all three protocols converged under one umbrella foundation. They have not. Each remains under its originating organization, with only A2A having moved to neutral Linux Foundation stewardship.
60 second bullets to scan on the way to the call.
What problem each of MCP, A2A, and ACP solves
Why the three are complementary layers, not competitors
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.