Model Context Protocol: what MCP actually fixed

September 13, 2026 · 5 min · Free, with full transcript

Anthropic released the Model Context Protocol as an open standard in December 2025. Was that generosity or strategy? The position here is that it was strategy that worked: Anthropic set the terms the ecosystem now builds on, and MCP solved the N times M integration problem by replacing bespoke connectors with one protocol. But fixing the plumbing exposed the reasoning failures that broken pipes used to hide. You get a way to read vendor numbers, including OpenAI's 70 percent task completion claim and Google's tool use accuracy figures, against MIT Technology Review's independent evaluations closer to half. You also get Block's approach of exposing a narrow, well documented tool surface, plus a rule for scoping what an agent may do without a human.

Listen to the podcast

5 min

Chapters

Key takeaways

  • Before connecting an agent, write down every action it may take without a human and cut anything irreversible.
  • Treat vendor benchmarks like OpenAI's 70 percent completion rate as directional only, and check independent evaluations such as MIT Technology Review's.
  • Expose a limited, well documented set of internal tools rather than everything at once, as Block did with its MCP agents.
  • Let agents draft refunds, pull data and format reports, but keep a human on the trigger for irreversible steps.
  • Expect composability: you describe the goal and the agent assembles the path, so you give up knowing the steps in advance.
Read the full transcript

Host:Welcome to the Leaders Insights podcast. Today's episode, the quiet handoff that changed what AI agents can actually do. In December 2025, Anthropic gave away infrastructure it could have charged a fortune for. Nine months later, was that generosity or was that strategy?

Expert:It was strategy dressed as generosity, and it worked. The model context protocol, MCP, basically a universal adapter that lets an AI agent talk to your tools, your files, your databases without custom glue code for each one, became an open standard. Anthropic didn't lose control by giving it away. They set the terms everyone else now builds on.

Host:That sounds like the story every vendor tells about, open. Why should I believe it changed anything real?

Expert:As before MCP, connecting an agent to say your company's sales force and your internal wiki meant two bespoke integrations that broke every time an interface changed. Engineers called it the end times M problem. Every model times every tool, a maintenance nightmare. MCP collapses that into one protocol, one socket, many plugs.

Host:Fine. But everyone's screaming that 2026 is the year of autonomous agents doing your job. What's the consensus getting wrong?

Expert:The consensus thinks the bottleneck was intelligence. It wasn't. The models were smart enough two years ago. The bottleneck was plumbing. Agents couldn't reliably reach the systems where actual work lives. MCP fixed the plumbing, not the brain. And here's the uncomfortable part. Better plumbing exposes how bad the brain still is at judgment.

Host:Meaning what, exactly?

Expert:Meaning an agent can now flawlessly connect to your billing system and confidently issue the wrong refund to the wrong customer at scale. Speed of access went up. Quality of decisions did not. O'Reilly's radar reporting through 2026 keeps landing on the same thing. The failures aren't connection failures anymore. They're reasoning failures that used to be hidden behind broken pipes.

Host:The labs would push back. OpenAI and Google both publish adoption numbers that look spectacular.

Expert:They do. And cross-check every one of them. But OpenAI says agent task completion jumped past 70% on their internal benchmarks. Remember, they sell the agents. That's a vendor grading its own homework. Same with Google AI's figures on tool use accuracy. Useful directionally. Worthless as gospel. MIT Technology Review ran independent evaluations this year and found real-world completion rates on multi-step business tasks sitting far lower, closer to half. Once you count the silent errors, nobody catches until an invoice is wrong.

Host:So companies rolling out agents right now are buying a faster way to make mistakes?

Expert:If they skip the boring part, yes. The boring part is scoping. The teams winning in 2026 aren't the ones giving agents the most autonomy. They're the ones drawing the tightest fences. Let the agent draft the refund, pull the data, format the report. Keep a human on the trigger for anything irreversible. It's a golden retriever with a company credit card. Enthusiastic. Capable. Absolutely not allowed to check out alone.

Host:But isn't that just old-fashioned automation with a chatbot on top? Where's the actual shift?

Expert:The shift is composability. The agent can now chain tools it was never explicitly programmed to combine. Pre-MCP, you scripted every step. Post-MCP, you described the goal. And the agent assembles the path from whatever tools are exposed to it. It's genuinely new. It's also exactly why the fencing matters. You gave up knowing the precise steps in advance.

Host:Give me a company that got this right.

Expert:Block, the payments company, formerly Square, built internal agents on MCP early and were public about it. Their move was to expose a limited, well-documented set of internal tools rather than everything at once. Narrow surface, clear permissions. Their engineers reported fewer runaway behaviors, precisely because the agent couldn't reach into systems it had no business touching.

Host:So the quiet handoff wasn't about capability. It was about who controls the standard.

Expert:Right. Anthropic gave away the protocol to own the ecosystem's shape. Smart. But for anyone actually deploying, the lesson is smaller and more useful.

Host:Which is?

Expert:Before you connect a single agent this quarter, write down the list of actions it may take without a human. If that list includes anything you can't undo, cut it. That one boundary saves you more than any model upgrade will.

Host:Sources for today's episode. The decoder, TechCrunch AI, Ars Technica AI, O'Reilly Radar, OpenAI, Vendor, AI Lab, KNuggets. That's your five. The full AI track, structure, no fluff, is at MBA-training.com.

Read and go deeper