The layer cake

Think of agent infrastructure as a layer cake. The model sits at the top. Underneath, the model calls tools (MCP). Agents talk to other agents (A2A). Agents live in a mesh with discovery, identity and policy (AGNTCY). Agents commit money on behalf of users (GeraNexus).

MCP

What it does: a standard for exposing tools to LLMs. Open, implemented by every major lab.

What it does not do: negotiate price, hold money in escrow, sign binding commitments, produce audit-grade receipts.

A2A (Agent-to-Agent)

What it does: a messaging and negotiation protocol for agents talking to each other across org boundaries. Shipped by Google with enterprise partners.

What it does not do: settle to a payment rail, bind a human consent scope, standardise dispute resolution.

AGNTCY

What it does: an open agent-mesh initiative (Cisco, LangChain, Galileo and others) focused on discovery, identity, policy and mesh-level plumbing.

What it does not do: define the transactional verb surface or provide an escrow primitive.

GeraNexus

What it does: four transactional verbs — intent, commit, settle, receipt — usable across payment rails and sitting on top of MCP/A2A/AGNTCY.

What it does not do: replace MCP for reads, replace A2A for peer negotiation, provide the mesh plumbing that AGNTCY provides.

Can you use them together?

Yes, and we think most serious deployments will. MCP as the read-side. AGNTCY for mesh-level discovery and identity. GeraNexus whenever money commits. A2A for cross-org peer negotiation.

Where comparisons go wrong

We see “MCP vs GeraNexus” headlines. They are not rivals. The right framing is layers. See also the direct MCP vs A2A comparison and the Agora/Skyfire/NANDA comparison.