What is ACP and how does it position relative to MCP and A2A in the 2026 agent protocol landscape?
Describe ACP (Agent Communication Protocol) and explain how it relates to MCP and A2A in the agent protocol ecosystem as of 2026. How is each protocol governed?
Three protocols, three boundaries: MCP wires agents to tools, A2A delegates tasks across orgs, ACP is REST-based general agent messaging. Complementary layers, no single governance body.
Think of a busy office. MCP is the standard wall socket every gadget plugs into to get power and data; it connects one worker to all their tools. A2A is the courier service that hands a sealed work order to another company's office and waits for the finished result. ACP is the open inbox where any worker can drop a rich message, attach files, and have a back and forth conversation, even if the two sides never agreed on a format in advance. None of these replaces the others. You still need sockets for your tools, couriers for outside jobs, and an inbox for general chatter. The catch in 2026: each system was built by a different landlord, so there is no single building manager making them all work together perfectly yet.
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 separates candidates who have skimmed headlines from those who actually track the agent-protocol space. The trap is to treat MCP, A2A, and ACP as three runners in one race, where one must eventually win. That framing is wrong, and a good interviewer is listening for whether you can place each protocol at the boundary it solves rather than ranking them against each other.
The cleaner mental model is that an agent system has multiple boundaries. There is the boundary between an agent and its tools or data. There is the boundary between one agent and another agent, especially across organizations. And there is the looser boundary of general agent to agent conversation that is richer than a single task handoff. Each protocol grew up to serve one of these, which is why they coexist instead of colliding.
This deep dive maps each protocol to a boundary, explains why MCP became the clear winner at the tool layer, examines the genuine overlap and convergence pressure between A2A and ACP, and then tackles the governance reality that makes this a senior-level question. Throughout, the honest stance is that the tool layer is settled while the agent-messaging layer is not. A strong answer holds that nuance instead of declaring a premature winner.
MCP: the settled tool-connectivity layer
MCP is the Model Context Protocol, released by Anthropic in November 2024. It standardizes the boundary between an LLM host application and external capability providers. A server exposes three primitives: tools (callable functions), resources (read-only data referenced by URI), and prompts (reusable templates). The host discovers these over JSON-RPC and surfaces them to the model, typically by registering each tool through the model vendor's function-calling format.
The problem it solved was the M×N integration explosion. Before MCP, every agent app hand-wrote a bespoke integration for every tool, so M apps times N tools meant M×N pieces of glue. MCP collapses that to M+N: write one server per tool, one client per host, and any host talks to any server. That reframing is the single most important sentence to be able to say in an interview, because it explains the protocol's entire reason for existing.
It is worth being precise about what MCP is not. It is not a way for one agent to delegate work to another agent, and it is not a model-side calling convention. It sits below the model's function-calling format and above the concrete tool, acting as the integration plane.
By 2026, MCP is effectively the default at the tool layer. It is an open spec, so although Anthropic created it, any provider can implement it. That openness plus an early ecosystem is exactly why it spread beyond a single vendor and why competitors did not displace it.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Dimension | MCP | A2A | ACP |
|---|---|---|---|
| Boundary | Agent to tools and data | Agent to agent, cross-org delegation | Agent to agent, general messaging |
| Origin | Anthropic, Nov 2024 | Google, April 2025 | IBM Research, BeeAI |
| Core unit | Tools, resources, prompts | Capability card, task lifecycle | Rich multi-turn messages |
| Transport style | JSON-RPC over stdio or HTTP | HTTP-based task exchange | REST-based messaging |
| Governance | Open spec under Anthropic | Linux Foundation | IBM and community |
| 2026 status | Dominant at tool layer | Leading cross-org delegation | Exploratory, overlaps A2A |
Real products, models, and research that use this idea.
- Anthropic maintains MCP as an open spec with reference servers; Cursor and Zed adopted it so one server runs across hosts.
- Google introduced A2A in April 2025 with launch partners, then donated it to the Linux Foundation for neutral governance.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy did MCP win the tool-connectivity layer when other tool standards existed?
Point at the M×N integration pain it targeted, early reference servers and SDKs, fast host adoption by Cursor and Zed, and the resulting network effects.
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.
Framing the three as competitors fighting for one slot. They sit at different boundaries: tool connectivity, cross-org delegation, and general agent messaging.
60 second bullets to scan on the way to the call.
The boundary each protocol owns: tools, delegation, general messaging
Origin org and rough date for MCP, A2A, and ACP
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.