Skip to content
AI Engineer Guide.

concept · concepts · intermediate · concept 14 of 16

A2A vs MCP

0 of 16 complete

Why it matters

Both protocols appear in the same architecture diagrams, and teams ask which one to standardise on. The distinction is not technical capability but trust and knowledge boundaries. An MCP server is a capability the host has chosen to load, with descriptions the host can read and filter. An A2A peer is a separate system with its own policy, and the caller is not entitled to its internals. Mixing the two models produces either over-sharing or an integration that cannot delegate.

Mental model

MCP is a kitchen with labelled drawers; A2A is ordering from another restaurant. The metaphor breaks because the other restaurant may ask you questions mid-order and may use its own suppliers, which is exactly the point: you do not get to inspect its kitchen.

How it works

Delegate across a boundary, compose inside itInteractive

A2A and MCP at two boundaries. Caller to Peer is A2A 1.0, and SendStreamingMessage opens the SSE stream; Peer to MCP server is the peer's private integration and is not visible to the caller.

Caller

planner

Peer

owner

MCP server

inside

Step 1/4 · refuse
2 · 401 · audience mismatcherror · Peer → Caller
HTTP/1.1 401 Unauthorized
WWW-Authenticate: Bearer error="invalid_token"

{ "error": "invalid_token",
  "detail": "audience mismatch",
  "expected_aud": "a2a:billing",
  "got_aud": "mcp:issues" }

The peer authorises its own caller. A valid
signature for another boundary is not enough.

Step 1: refuse. Hover, tap or tab to any message to inspect its payload.

Figure. The caller must hold a token for the peer, not for the peer's MCP server: the first SendStreamingMessage is rejected on audience, and the retry opens the task stream. Internally the peer calls its own MCP server; that lifeline is invisible to the caller, which sees only a task and an artifact.

What to notice. The failure is a peer treated as a tool: a credential minted for an MCP server is presented to an A2A peer, which checks audience and subject rather than accepting any valid signature. Wrapping a controllable database as a peer, and the double-authorisation gap, are omitted. Token values, ids, the 401 body and the CSV row are illustrative.

Text version

Sequence across Caller, Peer and MCP server. Step 1, refuse: Caller to Peer SendStreamingMessage with JSON-RPC id 1, Authorization Bearer aud mcp:issues, and a message m-01 from ROLE_USER with text 'Reconcile March'; Peer to Caller HTTP 401 with error invalid_token, expected audience a2a:billing and got mcp:issues. No stream is opened. Step 2, delegate: Caller to Peer SendStreamingMessage again with JSON-RPC id 2, the same message id and a token aud a2a:billing; Peer to Caller the first SSE event, whose data carries jsonrpc 2.0, id 2 and a task t-8842, context c-17, status TASK_STATE_SUBMITTED. The stream is open. Step 3, compose: an internal Peer to MCP server tools/call invoice_lookup with arguments invoiceId inv_221 and _meta naming protocolVersion 2026-07-28; MCP server to Peer resultType complete with structuredContent status paid and amount_cents 12900. The caller has no arrow to the MCP server. Step 4, complete: Peer to Caller SSE events with jsonrpc 2.0 and id 2 carrying artifactUpdate a-17 march-reconciliation.csv and statusUpdate TASK_STATE_COMPLETED. The caller sees the task and the artifact, not the peer's tools, prompt or memory.

  1. Ask who owns the capability and whether the caller may see its implementation.
  2. If the caller configures and supervises the capability, expose it as an MCP server with tools the host can filter.
  3. If the caller delegates a goal to a system with its own policy, use A2A and treat the peer as opaque.
  4. Keep credentials separate per boundary, with tokens scoped to the server or the peer rather than shared.
  5. Do not tunnel A2A semantics through MCP tools, because the host will treat the peer as a capability it controls.
  6. Document the composition, since an A2A peer that uses MCP internally is invisible to the caller by design.

Trade-offs

ChoiceGood forCostFails when
MCP for everythingOne integration surfaceThe host assumes authority it should not haveThe other system has its own users and policy
A2A for everythingClean ownership boundaryHeavyweight for a simple toolThe capability is a database the host already controls
Both, clearly separatedRight tool per boundaryTwo protocols to operate and secureThe team is small and the topology is simple
Bespoke HTTP between agentsFull controlNo discovery or lifecycle standardPartners need to integrate
Trade-off mapInteractive

Good for

One integration surface

Cost

The host assumes authority it should not have

Fails when

The other system has its own users and policy

  1. The other system has its own users and policy
  2. The capability is a database the host already controls
  3. The team is small and the topology is simple
  4. Partners need to integrate

Figure. Each choice with what it buys, what it costs, and the case where it loses.

What to notice. The fourth column is the decision: if the losing case describes your situation, the choice is wrong.

Text version

Four trade-offs: MCP for everything is good for One integration surface, costs The host assumes authority it should not have, and fails when The other system has its own users and policy. A2A for everything is good for Clean ownership boundary, costs Heavyweight for a simple tool, and fails when The capability is a database the host already controls. Both, clearly separated is good for Right tool per boundary, costs Two protocols to operate and secure, and fails when The team is small and the topology is simple. Bespoke HTTP between agents is good for Full control, costs No discovery or lifecycle standard, and fails when Partners need to integrate

Failure modes

  • Peer treated as a tool. The host passes its own credentials to an A2A peer and expects tool-level control. Debug: check the token audience and confirm the peer authorises the caller's subject, not the host's service identity.
  • Capability treated as a peer. A database is wrapped as an A2A agent and the caller loses the ability to filter its tools. Debug: compare the exposed capability list with the caller's permission set.
  • Double authorisation gap. Both sides assume the other checks access. Debug: attempt a request with a valid token but an unauthorised subject and see which layer rejects it.

In production

Monitor each boundary separately: MCP server health and tool errors on one side, A2A task states and peer availability on the other. Keep the credentials and incident runbooks separate too.

Common misconceptions

People think: A2A and MCP compete for the same job.

Actually: They connect different things: capabilities to a host, and agents to agents.

People think: Using both doubles the work.

Actually: The boundaries differ, so each protocol replaces bespoke code that would otherwise exist at that boundary.

Interview angle

A strong answer decides by ownership and opacity, then shows a composition where one agent uses MCP internally and exposes A2A externally. It also names the credential separation. Weak answers compare feature lists without saying who controls the capability.

Sources