A2A and MCP in a multi-agent design: when a mid-market company needs both

This guide is for the CTO or engineering lead deciding whether a second agent needs to talk to the first, not only to systems, and for the IT governance owner weighing whether the extra protocol is worth the audit it adds. Our MCP servers guide designs the agent-to-tool layer this page assumes already exists; our agentic AI architecture guide places the human checkpoint inside one agent's own loop. This page is the layer above both: what changes once a second agent enters the picture.

In this guide

• Do MCP and A2A solve the same problem?
• The two-layer pattern
• Does a mid-market company need A2A yet?
• A worked case: order-to-cash as a multi-agent design
• What changes once a second agent is the caller
• Where AI fits in a multi-agent design
• What delivery record exists, and what does not
• Alternatives: how else a company coordinates more than one agent
• Limits of this guide
• Frequently asked questions
• Next step

Do MCP and A2A solve the same problem?

No. MCP (Model Context Protocol) connects one agent to the tools, data and systems it needs to do its job: an order-status lookup, a CRM update, a document search. A2A (Agent2Agent) connects one agent to another agent: it lets an agent discover what a second agent can do, send it a task, and track that task to completion, as a protocol between two reasoning systems rather than between an agent and a data source. A company with one agent and five tools needs only MCP; nothing in that design discovers or delegates to another agent.

The two-layer pattern

Most companies using both protocols put A2A at the top, coordinating which specialised agent handles which part of a job, and MCP underneath each agent, connecting it to the systems that part needs. Remove either layer and a different part breaks: without A2A, agents cannot hand off work in a structured way; without MCP, an agent that received a delegated task has no governed way to act on it.

Does a mid-market company need A2A yet?

Most mid-market companies with one agent and several tools do not need A2A yet, because nothing in a single-agent design delegates work to a second reasoning system; MCP alone, scoped and governed, is the simpler answer. The question is not "how many tools does the agent use" but "how many separate agents, built or owned by different teams, does a task pass through before it is done."

A useful test: can every tool call already be traced to the one agent, and the one person, who approved it? In a single-agent design the answer is always yes, because one agent holds the whole task. The moment a task splits across two agents — a triage agent that classifies a request and a billing agent that acts on it — the trace has to follow the task across agents too, or an approved action loses the context of which agent actually asked for it. That is the point at which A2A earns its place: not when a second tool is added, but when a second agent is.

A worked case: order-to-cash as a multi-agent design

Our order-to-cash guide maps sub-steps of order-to-cash (order intake, credit check, invoicing, collections follow-up) to where an agent fits and where a person decides. A company that builds one agent per sub-step, rather than one large agent for the whole process, hits exactly this condition: a collections agent must hand a disputed invoice to a billing agent, or an intake agent needs a credit-check agent's answer first. Each agent still needs its own MCP connections to its own systems; A2A lets them discover each other and pass the task along with enough context to audit the handoff.

What delivery record exists, and what does not

Netbase has built the kind of systems a multi-agent design's agents would each need their own governed access to, and has designed the single-agent control loop and tool-exposure layer this page builds on:

• Multi-tenant cloud ERP for a US client (not named). Since 2020, offshore development and managing partner on a cloud ERP delivered as SaaS; phase 1 covers CRM, a knowledge base, work and project management and API integrations, the kind of system a specialised agent's own MCP connection would wrap.
• ERP and CRM consultancy (not named). Odoo consultancy and customisation, e-commerce-to-ERP integrations synchronising catalogues, inventory, orders and customer data by API, and a Vtiger-to-ERP upgrade; each is a defined, bounded action an individual agent's MCP tool would call, not a raw database connection.
• Delivered AI beyond these, anonymised: retrieval-augmented knowledge assistants, document AI and MLOps pipelines for clients not named here (F072).

What does not exist. No published Netbase record describes a delivered multi-agent system using the A2A protocol, an agent delegating to another agent in production, or a measured result from combining A2A with MCP. The decision rule, the two-layer pattern and the sequencing steps above come from Netbase's ERP/CRM integration and single-agent delivery practice (F070, F072, F076, F077) and the cited protocol sources, not one measured multi-agent project; no duration, cost, accuracy or adoption figure is claimed for any multi-agent or A2A deployment.

Alternatives: how else a company coordinates more than one agent

When not to add A2A yet. A single agent with several tools and no second reasoning system to hand work to does not need a delegation protocol; adding one is complexity with nothing yet to coordinate. Add A2A once a second specialised agent exists or is firmly planned.

In Netbase custom development the client owns the IP created for it, and most projects run on fixed-price contracts agreed after discovery. The agent-to-tool layer sits in systems and API integration and ERP and CRM development, the agent-facing design in AI integration, delivered through agentic AI automation and the enterprise AI agent platform, for sectors such as professional services.

Limits of this guide

• Both protocols are young and revise often; check the current specification before committing an architecture to either.
• The decision rule and two-layer pattern are a starting design, not a guarantee that a given multi-agent system is safe or correctly scoped.
• No duration, cost, accuracy or adoption figure is claimed for any multi-agent or A2A deployment.
• Adding A2A does not remove the need for the single-agent controls (approval, trace, kill switch) this cluster's other guides cover; it adds a layer on top, not a replacement.

Next step

Tell us how many agents your design actually needs and which systems each one has to reach, and we will book a solution review to scope the delegation layer, the tool connections and the approval trail across them. You can also see the agentic AI automation service or browse more Netbase insights.

Related services and solutions