Anthropic Staff Engineer: MCP + Skills Is the Real Architecture for 2026 Agents

While the community debates whether MCP is dead, an Anthropic Staff Engineer offers a different verdict: MCP isn’t obsolete, it just needs to be paired with Skills. AI agents in 2026 won’t stop at coding—they’ll span multiple SaaS applications to handle knowledge workers’ daily tasks.

What MCP Solves

The Model Context Protocol (MCP) provides a standardized way for AI agents to discover and invoke external tools. It’s essentially the agent’s “senses and hands”—telling the model what it can do and how to do it.

But tool discovery alone isn’t enough. Agents also need to know when to use which tool and in what order to combine them. That’s what Skills solves.

The MCP + Skills Combination

In the Anthropic engineer’s architectural description, the division of labor is clear:

• MCP: Tool discovery, protocol standardization, cross-platform compatibility. Through MCP, agents know what tools are available and what the input/output formats are.
• Skills: Capability encapsulation, scenario orchestration, permission control. Each Skill encapsulates the calling logic of a set of MCP tools, corresponding to a specific work scenario.

This layered design addresses a core challenge in agent development: as the number of tools grows, the model’s choice space explodes, leading to degraded call quality. Skills acts as an intermediate layer, organizing “raw tools” into “ready-to-use capabilities,” reducing the model’s decision burden.

From Coding Agents to Knowledge Worker Agents

The article makes a clear judgment: 2025 was the year of Coding Agents, 2026 is the year of knowledge worker agents that collaborate across SaaS applications.

This shift has several implications:

Dimension	2025 (Coding Agent)	2026 (Knowledge Worker Agent)
Primary tools	File read/write, code execution, Git	CRM, email, calendar, documents, databases
Task characteristics	Single-system closed loop	Cross-system collaboration
Success criteria	Does the code compile?	Is the business process fully executed?
Core challenge	Code quality and context management	Tool selection, permissions, error recovery

MCP’s standardization lets agents connect to various SaaS services, while the Skills layer orchestrates these services into concrete business flows.

Keys to Enterprise Agent Adoption

For teams evaluating agent solutions, several points are worth noting:

1. Don’t choose just MCP or just Skills. A single tool discovery layer isn’t intelligent enough; a single capability encapsulation lacks extensibility. The combination is what makes production-grade agents.
2. Define Skills based on business scenarios, not tools. First clarify what work the agent needs to do, then encapsulate the corresponding Skill, and finally use MCP to connect to underlying tools.
3. Permissions and auditing are essential for cross-SaaS agents. When agents can operate CRMs, send emails, and modify calendars, permission boundaries and operation logs matter more than model capability.

Quick Start

To experience MCP in Claude:

# Install Claude Code (if not already)
npm install -g @anthropic-ai/claude-code

# Add MCP configuration in your project
# Edit .claude/mcp.json
{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed"]
    }
  }
}

# Launch and use configured tools in Claude Code
claude

For Skills configuration, refer to Anthropic’s official documentation on Claude Skills.

What to Watch

Whether the MCP + Skills architecture becomes the industry standard depends on several variables: whether OpenAI’s function calling ecosystem will adopt a similar design, how Google’s BYO MCP will interface with the Skills concept, and whether enterprise agent frameworks like LangGraph and CrewAI will natively support this layering.

As more SaaS vendors release official MCP Servers, the agent’s accessible tool library will expand rapidly. The quality of Skills layer design will become the key differentiator between “works” and “works well.”

---

Primary sources:

• Anthropic Engineering: Building agents that reach production with MCP
• Anthropic Staff Engineer technical sharing on MCP and Skills