AI Agent Protocols Explained: MCP vs A2A vs SLIM vs ACP

Modern AI applications need two kinds of interoperability. First, a host needs a consistent way to reach tools such as GitHub, Slack, a filesystem, or Postgres. Second, one application may need to delegate a task to an independently deployed agent without knowing that agent’s model or framework.

The protocols fit into three layers. MCP is the capability layer: host → tool or resource. A2A is the agent layer: client → remote agent; ACP belongs to this same lane but is now mainly a migration concern. SLIM is an optional transport layer beneath them, providing secure routing and messaging when direct connections are not enough.

Running example: a support host looks up incident INC-42 through MCP, asks a remote research agent for analysis through A2A, and uses SLIM only if that A2A exchange must cross sites with end-to-end encryption and registration-based reachability.

What it is. Model Context Protocol (MCP), introduced by Anthropic, standardizes how an AI host connects to servers that expose tools, resources, and prompts. A client inside the host communicates with an MCP server, commonly through JSON-RPC over stdio or Streamable HTTP. The server then calls the real system—such as GitHub, Slack, a filesystem, or Postgres—and returns structured content.

How it works. Suppose a user asks, “Summarize incident INC-42.” The host discovers a lookup_issue tool, sends a typed call through its MCP client, and receives either incident data or a structured error. The model never receives database credentials and never talks directly to Postgres; the server owns that boundary.

Use MCP when several compatible hosts need reusable access to the same capability. Skip it when a local typed function is enough or when the destination is an independent agent with its own task lifecycle—that is an A2A problem.

What it is. A2A (Agent2Agent Protocol), introduced by Google and now governed through the Linux Foundation, standardizes communication with an independently deployed agent. The agents may use different models, frameworks, tools, and memory systems because A2A defines the external contract rather than their internal implementation.

How it works. An Agent Card describes the remote agent’s identity, skills, interfaces, and security requirements. A client discovers that card, sends a message or task, and receives task state, clarification requests, streamed updates, or final artifacts. A registry is optional: cards may come from a well-known URL, direct configuration, or a curated registry.

Example and failure: the support host sends INC-42 to a research agent. If the response times out after the remote agent accepted the task, blindly resending can create two investigations and two artifacts. Reuse a stable task ID, query task state before retrying, and mark one terminal result authoritative. Use A2A when the remote worker has a separate deployment and task lifecycle; use a function call when it does not.

What it is. SLIM means Secure Low-Latency Interactive Messaging. It is an open-source AGNTCY messaging framework—not an OpenAI tool-calling protocol. SLIM provides routing nodes, registration-based reachability, native streaming RPC (SRPC), and end-to-end payload encryption through Message Layer Security (MLS).

How it fits. A2A or MCP defines the meaning of a task or tool message; SLIM can carry that message across a distributed network. An endpoint registers with the SLIM network, sends encrypted content through one or more routing nodes, and the destination session decrypts it. Routing infrastructure can forward the payload without reading it, even where TLS terminates between hops.

Use SLIM when agents cross organizational or network boundaries and need secure multi-hop routing, group messaging, or reachability without exposing every endpoint. Skip it when direct HTTPS to a known agent already satisfies the threat model.

What it is. Agent Communication Protocol (ACP) was introduced by IBM’s BeeAI community as an HTTP-native contract for discovering agents and running work synchronously or asynchronously. Agent manifests described capabilities and content types, while REST endpoints carried messages and run state.

What changed. ACP addressed much of the same remote-agent boundary as A2A. The ACP project subsequently joined A2A under the Linux Foundation and now provides migration guidance. That makes ACP important for understanding or maintaining existing BeeAI services, but it is no longer an equal greenfield choice beside A2A.

Failure and recovery: a partial migration can map discovery but lose ACP run status or content types, leaving clients polling forever or dropping an artifact. Replay captured ACP success, waiting, failure, and cancel cases against the A2A adapter; keep the old endpoint available until every case reaches the same terminal state. Use ACP only when compatibility with an installed service requires it.

Choose the protocol by asking what the connection crosses. Choose MCP when an AI host needs tools, APIs, files, or data. Choose A2A when a client delegates work to an independently deployed agent. Evaluate SLIM only when those messages need secure multi-hop transport, registration-based reachability, or group communication. Keep ACP for an existing BeeAI integration while planning its A2A migration.

A real system may combine them. The support host can query Postgres through MCP, delegate investigation through A2A, and carry the A2A exchange over SLIM when the network requires it. But do not add every protocol to every hop: local modules should remain typed functions, and direct HTTPS is valid when one known endpoint and ordinary transport security are enough.

Official references (checked July 2026): MCP architecture and specification at modelcontextprotocol.io/docs/learn/architecture and modelcontextprotocol.io/specification; A2A documentation and specification at a2a-protocol.org/latest; ACP documentation and migration status at agentcommunicationprotocol.dev; SLIM documentation at slim.agntcy.org/latest. Versioned protocol guarantees belong to those specifications, not to this comparison.