Clarion Street

Protocol Engineering: The Standardization Wave Reshaping the Agentic Stack

·5 min read

Most writing about agent protocols stops at two names, MCP and A2A. I have covered both on this site: MCP’s stateless rewrite and A2A’s move into the Agentic AI Foundation. There is a third project in the same conversation that gets far less attention, and it is the one that just archived one of its own specifications.

Cisco open-sourced AGNTCY in March 2025 with LangChain and Galileo. The Linux Foundation took it on that July, with Cisco, Dell, Google Cloud, Oracle, and Red Hat as formative members and more than sixty-five supporting companies. AGNTCY calls its goal the “Internet of Agents,” and of the three it is the one that treats the directory as its product rather than a side feature.

What AGNTCY actually is

The Linux Foundation press release describes AGNTCY as open infrastructure for “discovery, identity, messaging, and observability among AI agents from different vendors and frameworks.” Four pillars, but the component set in the project documentation is wider.

  • The Agent Directory Service, a federated registry for publishing, verifying, and discovering agents.
  • OASF, the Open Agent Schema Framework, an extensible data model for describing an agent’s attributes, skills, and capabilities across A2A, MCP, and more.
  • SLIM, Secure Low-Latency Interactive Messaging, a network-level messaging layer with pub/sub, streaming, and MLS encryption.
  • Identity, decentralized identifiers and verifiable credentials for agents and tools.
  • SHADI, a hardened runtime for an agent with verified identity, gated secrets, and OS sandboxing.
  • Observability and evaluation tooling, plus CSIT for integration testing and CoffeeAGNTCY as a reference implementation.

That list is self-description, so read the component names as marketing until someone tests them. The shape is the useful part. AGNTCY is a collection of components rather than a single protocol, and they all answer one question: once agents can talk to each other, how do you find them, describe them, and route work to them.

Not a third way to talk to an agent

The press release is explicit about where AGNTCY sits relative to MCP and A2A. AGNTCY does not compete with either. The release says the project makes “A2A agents and MCP servers discoverable through AGNTCY directories” and supports “message transport over the SLIM protocol.” MCP is how an agent reaches a tool. A2A is how an agent reaches another agent. AGNTCY is how either one gets found and reached. It sits above both rather than beside them.

The directory job is not uncontested. A2A already has the Agent Card, a way for one agent to advertise its capabilities to another. MCP has working groups for a registry and for server cards. Discovery exists in all three places. The difference is that for A2A and MCP, discovery is a feature bolted onto a protocol, while for AGNTCY the federated directory is the product.

The Agent Connect Protocol, archived

The Agntcy Collective maintained two interface specifications. OASF describes agents. The Agent Connect Protocol, ACP, was the other one: a REST interface, specified in OpenAPI, for invoking and configuring remote agents over an API. On April 11, 2026, the owner archived the acp-spec repository read-only. It sits there with 168 stars and 100 commits, frozen.

The repository gives no reason for the archive. I read the “why” myself, and it is fairly plain from the jobs the two specs did. ACP existed to invoke remote agents over an API. Invoking agents is the job A2A already owns, and AGNTCY explicitly interoperates with A2A. Archiving ACP reads as AGNTCY dropping its own “invoke” interface and deferring that job to A2A, while keeping the “describe” interface, OASF, for itself. The result is a cleaner division: AGNTCY handles describing agents, A2A handles invoking them.

I present that as interpretation. The archive notice is a single line with no rationale, but it is consistent with everything else the project publishes.

A different governance room

One detail is easy to miss. AGNTCY joined the Linux Foundation in July 2025, before the Agentic AI Foundation existed. AAIF formed in December 2025 and now hosts MCP and A2A, which I covered when A2A moved. Of the three, AGNTCY is the one that does not sit under AAIF. It lives in the Linux Foundation proper. The “describe and route” layer and the “invoke” layer are governed in two different rooms that now have to coordinate.

Caveats

The archive is real and dated. Most of the rest is direction rather than settled fact.

The “65+ supporting companies” figure comes from the Linux Foundation’s announcement and measures attention, not production use. The component descriptions come from AGNTCY’s own documentation, and I have read the docs and the press release rather than every spec repository, so details like SLIM’s quantum-safe claim and SHADI’s sandboxing are unverified.

There is also a naming collision worth flagging, because it will trip up readers of my earlier piece. The archived ACP here is the Agent Connect Protocol, an invocation interface. It is not the Agent Commerce Protocol, a separate ACP that OpenAI contributed around agent payments. Same initials, different project.

I am also not re-litigating the governance questions around agents here. The control plane and its security surface have their own post. This piece is about one project and one archive.

What it means

The tool interface and the agent interface both exist now, MCP for tools and A2A for agents. What is still being worked out is the layer above them, the directory and the schema for describing what is out there. AGNTCY is the project that names that directory as its product, and the ACP archive is the first sign of it trimming its own redundancy.

Sources

  1. Linux Foundation Welcomes the AGNTCY Project — The Linux Foundation
  2. AGNTCY Documentation — docs.agntcy.org
  3. agntcy/acp-spec: Agent Connect Protocol Specification — GitHub
  4. Model Context Protocol specification repository — GitHub
  5. A2A Protocol — a2a-protocol.org