MCP connects an agent to tools and data; A2A connects an agent to other agents. Different layers, complementary, not competing.
Picture a worker at a desk. MCP is the worker reaching for tools on the desk: a calculator, a filing cabinet, a printer. It is how one worker grabs and uses equipment. A2A is two workers talking across the office to split a job: 'You handle the invoices, I'll draft the report, send me yours when done.' One protocol is about a single agent picking up tools. The other is about agents coordinating and delegating work to peers. They never fight over the same job. A real system uses both at once: each agent uses MCP to touch its own tools, and the agents use A2A to hand tasks back and forth.
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.
The MCP versus A2A question shows up because both arrived in the same 2024-2025 wave of agent protocols, both talk about agents, and both move JSON over a network. Candidates who skim the headlines assume one must be replacing the other, or that the difference is a transport detail. Neither is true. The interview value of this question is precisely that it sorts people who pattern-matched on buzzwords from people who understand the architecture.
The right mental model is two orthogonal axes. MCP is the vertical axis: it standardizes how a single agent reaches down to tools, files, and data. A2A is the horizontal axis: it standardizes how agents reach across to each other to delegate and coordinate work. They cross different boundaries, so they compose rather than compete. Anthropic released MCP in November 2024; Google introduced A2A in 2025 with dozens of partners and stated up front that it was complementary to MCP.
This deep dive separates the two cleanly, shows what each protocol actually carries on the wire, walks through a concrete system that runs both at once, and ends with the traps that turn this easy question into a quiet rejection. By the end you should be able to draw the two-axis picture on a whiteboard and explain why a peer agent is not just another tool.
The two-axis mental model
Picture an agent sitting at the center of a cross. The vertical arm points down to its capabilities: a filesystem, a database, an API. The horizontal arm points sideways to other agents it can talk to. MCP owns the vertical arm. A2A owns the horizontal arm.
MCP is the agent to tool boundary. It answers a single question: how does this agent access the outside world? The host connects to MCP servers, discovers a catalog of tools, resources, and prompts, and invokes them. The scope is always one agent's own reach.
A2A is the agent to agent boundary. It answers a different question: how do two autonomous agents coordinate? One agent advertises its skills, another sends it a task, and they exchange messages and results. The scope is the relationship between peers, not the relationship between an agent and its equipment.
Because the two arms point in different directions, there is no overlap to fight over. Asking which one wins is like asking whether a USB port competes with a phone call. They do unrelated jobs.
The axes also explain why the relationship is composition, not hierarchy. An agent is simultaneously a consumer of tools on the vertical axis and a peer on the horizontal axis. Add a third agent and the vertical arms multiply independently of the horizontal links. That separation is the whole point: each protocol can evolve on its own without breaking the other, and a team can adopt one before the other. Many systems start tool-only with MCP and grow into A2A once a second agent appears.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Concern | MCP | A2A |
|---|---|---|
| Boundary | Agent to tools and data | Agent to peer agent |
| Core question | How does an agent access capabilities? | How do agents delegate and coordinate? |
| Unit of interaction | Tool call, resource read, prompt | Task with messages and artifacts |
| Discovery | tools/list, resources/list | Agent Card at well-known URL |
| Openness | Open spec, Anthropic origin | Open spec, Google origin |
| Relationship | Complementary, lower layer | Complementary, peer layer |
Real products, models, and research that use this idea.
- Google introduced A2A in 2025 with partners, positioning it as complementary to Anthropic's MCP rather than a replacement.
- An orchestrator agent delegates a research subtask to a specialist agent via A2A, and that specialist queries Postgres through an MCP server.
What an interviewer would ask next. Try answering before peeking at the approach.
QWalk me through a system where an agent uses both MCP and A2A in a single workflow.
An orchestrator delegates a subtask to a specialist via A2A; the specialist uses MCP to read files and call a database, then returns an artifact over A2A.
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 A2A as a newer or rival version of MCP. They span different boundaries, so a real stack runs both side by side.
60 second bullets to scan on the way to the call.
Which boundary MCP spans versus which boundary A2A spans
The core question each protocol answers
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.