Quick answer
Use A2A when separately deployed agents need to discover each other, exchange tasks and return results across frameworks. Agent Communication Protocol, the BeeAI-associated ACP, has joined A2A; Agent Client Protocol, also called ACP, instead connects editors to coding agents. For agents inside one application, native delegation or workflow calls may need neither.
Which ACP are we comparing with A2A?
Two protocols share the same abbreviation. Agent Communication Protocol was developed alongside BeeAI for agent interoperability over a RESTful API. Its introduction now says ACP is part of A2A under the Linux Foundation. Agent Client Protocol is a different project: it standardises communication between code editors and coding agents, including local processes over JSON-RPC and stdio. Treating those two ACPs as interchangeable produces the wrong integration plan.
What does A2A put on the wire?
The A2A specification defines Agent Cards for discovery and messages, tasks and artifacts for exchanges. A remote agent can expose capabilities and return work without exposing its internal tools or memory. The specification defines JSON-RPC, gRPC and HTTP+JSON bindings; an individual product may implement only some of them. A protocol checkbox is therefore the start of a compatibility check, not the end.
What is the current status of each protocol?
Verified November 17, 2026: the A2A repository marks v1.0.1 as its latest release, while the specification site's latest-released-version banner still points to 1.0.0. The archived Agent Communication Protocol repository marks v1.0.3 as its latest release. These are repository release labels, not interchangeable SDK version numbers. ACP's own introduction directs readers towards A2A; do not select it as a competing new standard without an existing compatibility requirement.
Agent Client Protocol remains a separate editor-to-agent specification. No version is asserted here because the reviewed introduction does not establish one. Its JSON-RPC connection model is not the BeeAI ACP REST API and is not an A2A task endpoint. Ask for the protocol's full name before discussing versions.
Which tools in our matrix document each protocol?
This is a fresh documentation review, not a change to the site's existing matrix marks. The table covers the eight tools in the main feature matrix. 'Not documented in the pages we reviewed' means no claim was verified in those sources, not that an implementation cannot exist.
| Tool | A2A evidence | ACP evidence |
|---|---|---|
| BAND (band.ai) | A2A adapter and gateway | Agent Client Protocol server and client adapter, for editors and coding-agent processes |
| LangGraph | LangSmith Agent Server: v1.0 JSON-RPC; selected v0.3 methods accepted | Not documented in the pages we reviewed |
| CrewAI | A2A client delegation and server configuration | Not documented in the pages we reviewed |
| Google ADK | Exposing and consuming guides; Go, Java and Kotlin experimental | Not documented in the pages we reviewed |
| Microsoft Agent Framework | Foundry Agent Service publishing: v1.0 GA, v0.3 preview; service-specific evidence | Not documented in the pages we reviewed |
| Agno | AgentOS A2A interface for agents and teams | Not documented in the pages we reviewed |
| OpenAI Agents SDK | Not documented in the pages we reviewed | Not documented in the pages we reviewed |
| n8n | Not documented in the pages we reviewed | Not documented in the pages we reviewed |
BAND's ACP documentation means Agent Client Protocol, not the BeeAI Agent Communication Protocol. Its A2A path and ACP path solve different connection problems. BAND's operating limitation remains a hosted dependency: Hosted (SaaS); connects agents running on any machine, including AWS, GCP or on-prem. Neither documented adapter makes the hosted collaboration service a self-operated framework.
How should a team choose a compatible implementation?
Start at the remote endpoint. LangSmith documents the v1.0 JSON-RPC binding and selected older method names; that is narrower than every transport in the A2A specification. ADK separately labels some language integrations experimental. Microsoft's cited evidence belongs to Foundry Agent Service, not proof that every open-source Agent Framework deployment exposes the same endpoint. Check the product, language and transport together.
- For remote agent delegation, verify both endpoints' A2A versions, binding and task behaviour.
- For an existing BeeAI ACP integration, use the ACP project's A2A direction to plan compatibility work.
- For editor-to-coding-agent integration, check Agent Client Protocol rather than Agent Communication Protocol.
- Do not confuse an SDK package release with the protocol version it implements.
When does a team need neither?
When agents live inside one application, a subagent call, handoff or explicit workflow can carry the work without another wire protocol. LangChain documents those native patterns. A protocol becomes useful at a separately deployed boundary, not simply because two agents exist. Shared collaboration is another requirement again: task exchange alone does not specify the room, participants or shared context that the application needs.