MCP, Agent Skills, and A2A: The Three Layers Every Enterprise AI Strategy Needs

The AI agent ecosystem has fractured into camps on social media, and almost all of them are wrong.

One side says MCP is all you need. Another insists Agent Skills have made MCP obsolete. A third crowd pushes Google's Agent-to-Agent protocol as the real future. I've been building with all three across enterprise implementations, and I want to say something plainly: these technologies solve different problems at different layers of the stack. Arguing about which one wins is like debating whether a phone number is better than a conversation is better than a conference call. They do different things.

Executive Read

If you only have a minute, here's what you need to know.

• MCP is a connectivity protocol. It standardizes how AI agents connect to external tools and data, the same way USB-C standardized device peripherals.
• Agent Skills are knowledge packages. They encode your organization's procedures, standards, and expertise into portable folders that agents can discover and apply.
• A2A is a collaboration protocol. It standardizes how AI agents discover each other and coordinate on tasks, regardless of vendor or framework.
• They operate at different layers: MCP handles "can the agent reach Salesforce?" Skills handle "does the agent know how we use Salesforce?" A2A handles "can the sales agent hand off a task to the finance agent?"
• All three are open standards. The organizations building on all three layers now will have a meaningful head start as the ecosystem solidifies.

Three Problems, Three Standards

Here's a scenario I've seen play out repeatedly. An enterprise connects an AI agent to their internal tools. The connection works. The agent can read tickets, pull records, query data. Then someone asks the agent to create a Jira ticket for a production incident, and it produces something that looks nothing like how their organization structures incident reports. And when they need the agent to coordinate with a compliance team's agent, there's no standard way to make the handoff happen.

Three different gaps. MCP solves connectivity: how do you give an agent standardized access to tools and data? Agent Skills solve knowledge: how do you teach an agent your specific procedures? A2A solves collaboration: how do your agents discover each other and coordinate on complex tasks?

MCP: The Connectivity Layer

The Model Context Protocol, introduced by Anthropic in November 2024 and donated to the Linux Foundation in December 2025, replaced the fragmented world of custom AI-to-tool integrations with a single protocol. An MCP client can connect to any MCP server, whether it exposes a database, a SaaS app, or a file system. The server describes its tools using a standard schema, the agent calls them using a standard interface.

The ecosystem has grown fast. Gartner projects 75% of API gateway vendors will have MCP features by 2026. Google, Microsoft, OpenAI, and AWS have all adopted the standard. At Block, employees reported saving 50-75% of their time on common tasks using MCP-powered tools.

An MCP server exposes three primitives: tools (operations the agent calls), resources (data it reads), and prompts (pre-built interaction templates). The March 2025 spec introduced Streamable HTTP for serverless deployment. In January 2026, MCP Apps added support for interactive UI components rendered directly in conversations.

The value is real. But MCP gives the agent the ability to act. It doesn't give the agent judgment about how to act within your organizational context, and it doesn't help your agents talk to each other.

Agent Skills: The Knowledge Layer

Agent Skills, published as an open standard at agentskills.io in December 2025, package organizational knowledge into portable, reusable components. A skill is a folder containing a SKILL.md file with instructions, optionally accompanied by scripts and reference documents. You can read the entire specification in a few minutes.

Skills solve the problem that's been quietly undermining enterprise AI for two years: LLMs are general-purpose by design, and your business processes are specific by necessity. When you ask an AI agent to review a pull request, a generic agent checks for bugs and style. It won't know your team requires a specific approval workflow for database migrations, or that changes to the payments module need additional security review. A skill encodes all of that.

The architecture uses progressive disclosure. Research from late 2024 found that LLM decision-making degrades when presented with more than 20-25 tools. Skills solve this by loading knowledge in layers: metadata at startup (~100 tokens per skill), full instructions when a matching request arrives, and resource files only on demand. The knowledge a skill can contain is effectively unlimited while context consumption stays bounded.

Microsoft adopted skills in VS Code and GitHub Copilot. OpenAI added support to ChatGPT and Codex CLI. Cursor, Goose, Amp, and OpenCode have integrated. Partner-built skills from Atlassian, Figma, Canva, Stripe, Notion, and Zapier are available.

A2A: The Collaboration Layer

Google's Agent-to-Agent protocol, announced in April 2025 and donated to the Linux Foundation in June 2025, addresses what happens when agents need to work together. Your sales agent needs to hand off a lead to onboarding. Your monitoring agent needs to tell your scaling agent about a traffic spike. Your compliance agent needs to review actions taken by procurement.

If MCP is a USB-C port connecting your laptop to peripherals, A2A is the network protocol connecting multiple computers to each other.

Three concepts worth understanding. Agent Cards: every A2A agent publishes a JSON file at /.well-known/agent.json listing capabilities, endpoint, and auth requirements, letting agents discover each other. Task Lifecycle: A2A introduces a Task object with defined states (submitted, working, input-required, completed, failed), enabling coordination on work that takes hours or days. Opacity: agents collaborate without sharing internal memory or proprietary logic. Your procurement agent works with a supplier's pricing agent without either revealing how they reach their decisions.

The protocol builds on familiar web standards: JSON-RPC 2.0 over HTTPS, Server-Sent Events for streaming, standard OAuth/OIDC for auth. Over 150 organizations support A2A, including Adobe, ServiceNow, Salesforce, and SAP. Tyson Foods and Gordon Food Service are using A2A agents in production supply chain workflows.

I want to be direct about maturity though. A2A is newer and less battle-tested than MCP. Developer adoption has lagged behind MCP's community momentum. Its value shows up most clearly at organizational scale where you genuinely have multiple agents from different vendors that need to coordinate. For most enterprises, this is a "plan for it now, implement it soon" technology.

How They Stack Together

The real power emerges when you combine them. Take automating a quarterly financial close:

MCP connects agents to your ERP, data warehouse, and reporting tools. A skill tells the agent: pull revenue by region and product line, calculate growth using the CFO's approved methodology, flag regions deviating more than 15% from forecast, route through the three-stage approval process. A2A lets the reporting agent hand off to a compliance agent that checks regulatory issues, which coordinates with an audit agent that verifies data integrity.

Without MCP, agents can't reach your data. Without skills, they don't know your procedures. Without A2A, they can't coordinate with each other.

Security Across the Stack

The security profiles differ in ways that matter.

MCP uses process isolation. Each server runs separately with its own credentials. The gap: over half of MCP servers still rely on static API keys rather than OAuth. That's a credential management problem worth taking seriously.

Skills execute within the agent's own process. More flexible, wider attack surface. In late January 2026, researchers found 341 malicious skills on a third-party marketplace using credential exfiltration and typosquatting. Govern skills like you govern code.

A2A has the strongest security model on paper: short-lived tokens scoped per task, signed agent cards, and opacity by default. The downside is implementation complexity. Enterprise SSO and mTLS across a multi-agent deployment isn't trivial.

Each layer needs its own security governance. Don't assume securing your MCP connections covers your skills and agent-to-agent communications.

The Social Media Debates Are Wrong

"Skills killed MCP." Works for individual developers using curl in a skill. Falls apart at enterprise scale when hundreds of developers need governed access to the same Salesforce instance. MCP provides centralized connection management. The real tension isn't MCP vs. Skills but MCP vs. traditional CLI methods. Skills just make those traditional methods easier to document.

"A2A is dead because MCP won." This conflates two problems. MCP won tool-connectivity. A2A addresses agent-to-agent collaboration, which MCP never targeted. The fair criticism: A2A over-specified too early and developer adoption lagged. That's a timing issue, not proof the problem doesn't exist. As enterprises deploy more specialized agents, the coordination problem will get louder.

Where to Start

Start with MCP if your agents are isolated from needed systems. Focus on the integrations teams request most.

Start with skills if your biggest gap is inconsistency. Identify your five most common AI-assisted workflows, interview the people who do them best, encode their approach.

Add A2A when you have multiple specialized agents that need to coordinate. Probably not your first move, but plan for it architecturally.

For measuring ROI: track time to complete standardized tasks, output consistency across team members, manual handoffs eliminated, and context switch frequency.

The Bigger Picture

The infrastructure layer for AI agents is standardizing in real time. MCP handles connectivity. Skills handle knowledge. A2A handles collaboration. None is going away. The organizations building across all three layers now will have a meaningful advantage as the ecosystem solidifies.

The window isn't closing tomorrow. But every month you wait is a month your competitors might be encoding their expertise into skills, standing up MCP connections, and wiring their agents to coordinate.

I'd rather be early than catching up.

References

1. Anthropic. "Introducing the Model Context Protocol." November 2024. anthropic.com
2. Anthropic. "Introducing Agent Skills." October 2025. Open standard at agentskills.io
3. Google. "Announcing the Agent2Agent Protocol (A2A)." April 2025. developers.googleblog.com
4. Block Engineering. "MCP in the Enterprise." April 2025.
5. Gartner. 2025 Software Engineering Survey.
6. Astrix Security. "State of MCP Server Security 2025." October 2025. astrix.security
7. Friedrichs-IT. "Agent Skills vs MCP: Two Standards, Two Security Models." February 2026. friedrichs-it.de
8. SmartScope. "The MCP vs Agent Skills Debate: Untangling the Category Error." 2025.
9. Google Cloud Blog. "Agent2Agent protocol is getting an upgrade." July 2025. cloud.google.com
10. Paramanayakam et al. "Less is More: Optimizing Function Calling for LLM Execution on Edge Devices." arXiv:2411.15399, 2024.

Connect with me on LinkedIn to discuss implementation strategies for your organization.

Matthew Kruczek

Managing Director at EY

Matthew leads EY's Microsoft domain within Digital Engineering, overseeing enterprise-scale AI and cloud-native software initiatives. A member of Microsoft's Inner Circle and Pluralsight author with 18 courses reaching 17M+ learners.