MCP Is Already Not Enough

Key Points

• We’re moving through three eras: APIs (humans wire things together), protocols like MCP (AI connects to tools), and agent-native networks (agents find and hire each other autonomously).
• MCP solved the “how does AI talk to apps” problem, but it still requires human setup, eats up to half the context window, and has serious security gaps.
• The real future isn’t better protocols. It’s agents that discover, negotiate with, and transact with other agents without anyone configuring anything.

Six months ago, MCP felt like magic. Connect your AI to Gmail, your calendar, your database, and suddenly one interface could do everything. Apps are becoming data feeds. MCP is the plumbing that makes it work.

But I’ve been building with MCP daily, and the cracks are showing. Not because MCP is bad. It’s genuinely useful. But because it’s a bridge technology. It solved the first problem (how does AI talk to apps?) while revealing a much bigger one: the current model still requires a human to set everything up, maintain it, and decide what connects to what.

That’s not the future. That’s a stepping stone.

Three eras, stacking fast

We’re moving through three distinct eras of how software communicates, and we’re barely into the second one.

Era 1: APIs. Where we’ve been for 15 years. Apps expose endpoints. Developers write code to connect them. Zapier made it easier, but every connection is still hand-built by a human.

Era 2: Protocols. Where we are now. MCP, Google’s A2A, China’s Agent Network Protocol. These standardize how AI talks to tools and how agents talk to each other. Massive improvement. But someone still has to find the right server, configure it, manage authentication, and troubleshoot when things break.

Era 3: Agent-native. What’s coming. Agents discover other agents and services on their own. They verify trust, negotiate capabilities, and transact without a human configuring anything. Your agent finds what it needs the same way you find a website: it searches, evaluates, and connects.

Each era doesn’t kill the previous one. APIs still exist. But the layer of interaction moves up, and the old layer becomes invisible infrastructure.

Why MCP isn’t the end state

MCP works. I use it every day. But its limitations point directly at what comes next.

It’s expensive on context. MCP tool descriptions consume 40-50% of the available context window before your AI does any actual work. Connect five or six services and your AI is half-full of instructions before it reads your first message.

Security is immature. 43% of MCP servers have flaws in their OAuth authentication. No standardized audit trail, no cost attribution, no governance framework.

It’s still human-configured. Every MCP connection requires a person to find the server, add it, configure API keys, and manage permissions. Your AI can’t decide it needs access to a new service and go get it. That’s the big one.

These aren’t fatal flaws. They’re growing pains. But they reveal that MCP is solving yesterday’s problem (standardizing connections) while the real problem is shifting (making connections unnecessary to manage at all).

What agent-native actually looks like

The Agent Network Protocol has an interesting idea: agents should be discoverable the way websites are. They publish descriptions of what they can do. Other agents search for them, verify their identity, and connect. No human in the loop.

Google’s A2A protocol tackles a different piece: agents delegating to other agents. Your personal agent needs legal research done, finds a specialized legal agent, negotiates the task, and gets results back. Agents hiring agents.

Combine these and you get something that looks less like a protocol stack and more like an economy. Agents with identities, reputations, and capabilities. Marketplaces where they find specialists. Trust frameworks where track records matter more than API keys.

This is already starting. Picsart launched an agent marketplace where creators “hire” AI assistants. Google Cloud Marketplace lets enterprises deploy pre-vetted agents. McKinsey has 25,000 agents working alongside 40,000 employees. Jensen Huang predicts 100 AI agents per human worker by 2036. This kind of leverage is already making the one-person company a reality.

MCP is the dial-up modem

If you’re building a product today, support MCP. But don’t mistake the protocol era for the destination. The products that win long-term will be discoverable by agents, not just humans. They’ll have semantic layers that explain what data means, not just schemas that describe what fields exist. Authentication will work agent-to-agent, not just human-to-service.

Every major technology shift has a bridge generation that feels revolutionary and looks primitive in retrospect. Dial-up until broadband. WAP browsing until the iPhone. MCP is that bridge. It proved AI can interact with external services through a standard protocol. That’s a big deal. But it inherited all the old assumptions: humans configure, humans manage, humans decide what connects to what.

The next era drops those assumptions entirely. And if you’re paying attention, you can already see MCP for what it is: not the future, but the proof of concept that made the future inevitable.