top of page

MCP, A2A, Agent Skills: A Plain-English Guide for Workflow People

Nov 4
3 min read

Why workflow people should care

The plumbing of AI agents just got a lot more official. In the past two weeks, the standards that connect agents to tools, to know-how and to each other moved under neutral governance and into the platforms where enterprise work actually runs.

You don't need to read a protocol spec to keep up. You need to know what each standard connects, because every connection is a handoff in a workflow. And every handoff needs an owner.

The standards at a glance

Three standards matter most for workflow design. Each one connects something different.

Standard

What it connects

In plain English

The workflow question it raises

MCP (Model Context Protocol)

An agent to tools and data

A universal adapter that lets an agent read from and act in your systems

Which systems can this agent touch, and who approved that access?

A2A (Agent2Agent)

An agent to other agents

A common language agents use to hand work to each other, even across vendors

When one agent passes work to another, who owns the result?

Agent Skills

Know-how to an agent

A packaged playbook of instructions, scripts and resources an agent loads when a task calls for it

Who writes the playbook, and who keeps it current?

A simple way to remember it: MCP is how an agent reaches your systems, A2A is how it reaches other agents, and Skills are how it learns your way of working.

What's changing

The standards now have a neutral home. On October 22, the Linux Foundation's Agentic AI Foundation added four projects, including Agent Skills, contributed by Anthropic. The foundation now hosts 10 projects, up from three when it launched in December 2025, including both MCP and A2A. The pitch, in its CTO's words, is a place to build shared standards "without any single company controlling the outcome."

The big platforms are wiring them in. Microsoft 365 Copilot now lets agents built in its toolkit connect to other agents over A2A, alongside existing MCP connections. Users can't add these agents themselves; an admin has to review and approve each one first. At its TechEd conference, SAP said its Joule agents will connect to outside systems and agents through MCP and A2A, and that third-party agents will be able to call Joule agents through a governed gateway.

The direction is clear. Agents from different vendors will increasingly hand work to each other inside the same business process.

The catch: cross-vendor workflows, cross-vendor gaps

Open standards solve a technical problem. They make it easy for agents to connect. They don't solve the workflow problem of who is responsible once they do.

Picture a common case. An agent in your productivity suite picks up a customer request, asks an agent in your ERP to check an order, and that agent calls a third-party agent to update shipping. Three vendors, three admin consoles, three sets of logs. If the customer gets the wrong answer, where do you look, and who fixes it?

The vendors are building controls for their own side of the handoff. Microsoft requires admin approval for each connected agent. SAP is routing agent traffic through a gateway that handles policy, monitoring and auditing. Those are useful, but each one governs the handoff from inside one vendor's walls. Nobody's console shows the whole workflow end to end.

That gap belongs to the workflow owner. It always has. The difference now is that the handoffs are faster, more numerous and harder to see.

What practitioners should do

  1. Map agent handoffs like any other handoff. On your workflow maps, show where an agent reaches a system (MCP), hands work to another agent (A2A) or relies on a packaged playbook (a skill). If it isn't on the map, nobody owns it.

  2. Name an owner for every cross-vendor handoff. Not the vendor and not "IT." A person accountable for that step's outcome, with the authority to pause it.

  3. Treat skills as process documentation. A skill is your way of working, written for an agent. It needs an author, a review cycle and version control, the same as any standard operating procedure.

  4. Get into the approval loop. Admin approval for connected agents is becoming standard. Make sure the person approving an agent connection understands the workflow it's joining, not just its security profile.

  5. Ask for the end-to-end view. When you evaluate tools, ask how you'll trace one piece of work across every agent and vendor it touches. Today the honest answer is often "you can't," and that's worth knowing before you build on it.

The standards will keep maturing. The question of who owns the work in between them is yours to answer now.

Sources

 
 
 

Comments


bottom of page