The integration problem that MCP solves
Every AI agent needs tools. It needs to read your CRM, query your database, send emails, update records. Before MCP, the only way to give an agent access to a tool was to write a custom integration for it — a bespoke function that called the tool's API in the specific way the agent expected. If you had ten tools, you wrote ten integrations. If you changed tools, you rewrote integrations.
Model Context Protocol (MCP) is Anthropic's answer to this. It's an open standard — think of it as USB-C for AI tools. An MCP server exposes a tool's capabilities in a standardised format. An MCP-compatible agent can connect to any MCP server. You write the integration once, on the tool side, and any agent that speaks MCP can use it.
Why this matters more than it sounds
The practical implication: an agent built on MCP today can connect to any new tool that ships an MCP server in the future — without you changing anything in the agent. As the MCP ecosystem grows (and it's growing fast — HubSpot, GitHub, Slack, Notion, and dozens of others already ship MCP servers), your agent's capability expands automatically.
Contrast this with a custom-integrated agent: every new tool connection requires a development sprint. The agent's capability is frozen at the point you last paid someone to extend it.
For UK businesses, the economics of this are significant. A well-architected MCP agent is a compounding asset. A custom-integrated agent is a depreciating one.
MCP Atlas and what the benchmarks show
Anthropic now publishes MCP-Atlas as a benchmark specifically for evaluating how well models use MCP tools in multi-step tasks. Claude Opus 4.8 scored 82.2% on MCP-Atlas (up from 77.3% on 4.7). The gap between Claude and other models on this specific benchmark is one of the larger ones in the current leaderboard — it reflects both the model's reasoning quality and the architectural advantage of being designed for MCP from the ground up.
The A2A dimension
MCP handles the tool connection problem. A2A (Agent-to-Agent protocol) handles the agent coordination problem — how agents delegate to each other and share state. These two protocols are complementary, not competing. An agent fleet where agents communicate via A2A and connect to tools via MCP is the architecture that will still be maintainable in three years.
We build all our agents MCP-ready from day one. Not because it's technically elegant — though it is — but because it's the only build decision that ages well.
What to ask any AI vendor
When a vendor shows you an AI agent demo, ask two questions:
1. Is it MCP-compatible? If not, ask how you add new tool integrations in 18 months.
2. How do you handle agent-to-agent coordination? If the answer is "it's all in one prompt," you've found the ceiling.
If they can't answer either question, the demo is a prototype, not a system.