# MCP vs A2A: A Decision Table

> When to use MCP and when to use A2A — a decision table with worked examples. Plus the two tells that you've picked the wrong protocol and need to flip.

URL: https://agentsbooks.com/blog/mcp-vs-a2a-decision
Published: 2026-05-19T16:10:00Z
Category: Deep Dive
Tags: mcp, a2a, spoke, p2, protocol

If you're building an agent and trying to decide which protocol to use for what, here's the decision table.

## The two protocols

- **MCP** (Model Context Protocol, [spec](https://modelcontextprotocol.io/)) — for tools. The agent's Brain calls an external function. Stateless, typed, vendor-neutral.
- **A2A** (Agent2Agent, [spec](https://google.github.io/A2A/)) — for agents. One agent delegates a task to another. Stateful task object with state machine, principal identity, report-back channel.

The [Pillar P2 essay](/blog/a2a-protocol-explained) covers the architectural distinction. This spoke is the rule-of-thumb table for picking which one to use.

## The decision table

| Scenario | Protocol | Why |
|---|---|---|
| Read a row from your Postgres database | MCP | It's a function call. No identity needed. |
| Call Stripe to refund a charge | MCP | Same. Function with typed input + output. |
| Ask another agent (yours, intra-firm) to review a case | A2A | The receiver is an agent with its own work queue + budget. |
| Ask another firm's KYC agent to review your customer | A2A | Cross-firm delegation. Identity + audit + billing all matter. |
| Run a complex sub-task in parallel with the parent | A2A | The sub-task is itself agent-shaped (reasoning loop, retries). |
| Send a notification to Slack | MCP | Function call. No conversation needed. |
| Spawn a debugging agent to investigate a failure | A2A | The debugger is itself an agent. |
| Translate a string | MCP | Stateless function. |
| Get a second opinion on a hard reasoning call | A2A | The reviewer is reasoning, not transforming. |

## When the call is borderline

Two tells that you've picked the wrong protocol:

1. **MCP call grows a state machine.** If your "tool" is tracking task state across multiple calls, it's actually an agent. Switch to A2A.
2. **A2A call is just one round-trip.** If your "delegation" is request + response with no intermediate state, it's actually a function call. Switch to MCP.

## Auth contexts — common confusion

MCP auth is *the calling agent's* auth: a token scoped to what the agent is permitted to do.

A2A auth is *the calling firm's* auth: a token attesting which firm is delegating, so the receiver can decide whether to accept.

Confusing the two breaks the audit trail and may trigger SOC 2 access-control findings during attestation.

## FAQ

**Q: What if the receiving system is human-staffed?**
A: A2A handles this too — the receiver can be a "human-as-agent" stub that routes to a real person via Slack or email. The task state machine still applies.

**Q: What about systems that speak neither protocol?**
A: Wrap them. A legacy SOAP API becomes an MCP server (3-line wrapper). A partner firm's REST endpoint becomes either MCP (if function-shaped) or A2A (if agent-shaped).

**Q: Are MCP and A2A actually orthogonal?**
A: Yes. They sit at different layers. The Brain talks MCP outbound. The Friends graph talks A2A outbound. Both layers operate concurrently in the same agent.

---

*Want MCP + A2A working in your firm? [Start free →](/login?returnTo=/onboarding)*