The query "a2a vs acp vs agntcy" carries the same buried premise as every "vs" search: three live contenders, one survivor, a bet you can lose. For most standards races that framing is fair. For this one it is already out of date, and the way it's out of date is the actual story. Two of the three answers were settled in 2025. The third isn't competing with the other two at all.

Start with the one that's still standing.

A2A is the application layer, and it won that layer#

A2A, Google's Agent2Agent protocol, is the part of the stack people actually mean by "agents talking to agents." It is a thin, framework-agnostic application protocol: agents publish a self-describing Agent Card — a JSON manifest of who they are, what they can do, and how to reach them — and then exchange tasks and messages over JSON-RPC 2.0 on plain HTTP, with SSE for streaming and push notifications for long jobs. Crucially, it treats the agent on the other end as a black box: you delegate a task, you don't get to see its tools or internal state.

The governance question that haunts protocols is also already answered. Google donated A2A to the Linux Foundation in June 2025, with AWS, Cisco, Microsoft, Salesforce, SAP, and ServiceNow as founding members. The repository carries roughly twenty-four thousand stars and SDKs across Python, Go, JavaScript, Java, and .NET. This is what a winning application-layer standard looks like: vendor-neutral, boring, widely implemented.

ACP didn't lose. It joined.#

The second contender, ACP, is where the "vs" framing breaks outright. IBM's Agent Communication Protocol — the one that powered the open-source BeeAI platform — was a real, well-designed alternative to A2A. Its pitch was that it spoke plain REST defined by OpenAPI instead of JSON-RPC, with first-class multimodal messages, sessions, and an await primitive that let an agent pause to ask for more information. For a few months in 2025 it was a genuine fork in the road.

Then it ended. On 27 August 2025, IBM archived the i-am-bee/acp repository read-only, with a notice that reads, in full cheerfulness, "ACP is now part of A2A under the Linux Foundation." Two days later the LF AI & Data community posted "ACP Joins Forces with A2A," IBM's incubation lead joined A2A's technical steering committee, and BeeAI's official posture became: A2A between agents, MCP for tools.

A real standards war does not end five months after launch. ACP's merger wasn't a defeat — it was the industry deciding, fast, that it had no appetite for a second one.

That timing is the load-bearing detail. MCP took the better part of a year and a public foundation fight to consolidate the tools layer. The agent-to-agent layer skipped the war entirely. Everyone had just watched what a protocol schism costs, and nobody wanted to spend another year of developer mindshare re-litigating JSON-RPC versus REST when the two designs were ninety percent the same idea. Consolidation, not competition, is the lesson the second generation of agent protocols learned from the first.

AGNTCY isn't racing A2A — it's the road A2A drives on#

The third name is the one that doesn't belong in the comparison the way the query implies. AGNTCY, launched by Cisco's Outshift incubator with LangChain and Galileo and donated to the Linux Foundation in July 2025, is not a rival application protocol. It's an infrastructure stack — the "Internet of Agents" — and it operates a layer down from where A2A lives:

You can see the relationship in AGNTCY's own repos, which include adapters that tunnel A2A and MCP over SLIM. It answers "how do agents find, trust, and securely route to each other," which A2A leaves almost entirely to you. Asking "A2A or AGNTCY" is like asking "HTTP or DNS": they're at different altitudes, and you use both.

The stack the query was actually about#

So the three-way fight collapses into a three-layer stack, which is the genuinely useful way to hold it. MCP is the tools layer — your agent reaching down to data and capabilities. A2A is the conversation layer — agents reaching sideways to delegate to peers. AGNTCY is the infrastructure layer — the directory, schema, and secure transport underneath both. ACP is a synonym for A2A you'll still see in older docs. (And the other ACP, the Agentic Commerce Protocol, is a payments standard with nothing to do with any of this — the acronym collision is real and unfortunate.)

The practical takeaway is short. You almost certainly want A2A plus MCP, and you reach for AGNTCY only when you actually have its problems: cross-organization discovery, a shared identity schema, a transport you can't otherwise harden. Picking a "winner" among the three was never the real decision, because the systems that fail in production rarely fail at the protocol — they fail at the layer you forgot you needed. Find your gap in the stack first. The protocol question mostly answers itself.