This week Google did something genuinely new: it let outside AI agents, Claude, ChatGPT, even Google's own Antigravity agent, connect straight into people's homes. Ask one of these agents to check your porch camera, nudge the thermostat down, or turn off a light you forgot, and it can now just do it. Google also drew one line it refused to cross. No agent, however capable, is allowed to unlock your front door. Not an option you can enable later. It's a rule built into the system itself.
That refusal is the actual story here, not the smart-home news. A company with more engineering resources than almost anyone on earth looked at what it had just built and decided one action was too consequential to hand over, full stop. If Google doesn't trust its own launch with that one thing, it's worth understanding why, because the same choice is coming to the tools you already use to run your business.
What actually changed this week
The mechanism behind this is called MCP, short for Model Context Protocol. Anthropic, the company behind Claude, built it as a shared standard for how an AI agent connects to outside systems. Before MCP, every AI tool needed its own custom-built plug for every other tool it wanted to use: a calendar app needed one connector, an accounting tool needed a different one, and nothing lined up. MCP is closer to a universal socket. Once a tool builds one door that speaks MCP, any AI agent that also speaks MCP can walk through it, read what's inside, and, if the door's owner allows, act on it directly.
That's the real shift. It's not that AI agents got smarter this month. It's that they got a standard way to reach into the software you already run, not just describe what you should click.
What this looks like in an ordinary business
Picture a dry cleaner who uses an online system to manage drop-offs and pickups. At some point, that same software will likely offer to "connect your AI assistant," the same pitch Google is making for houses now. Done well, that's genuinely useful: the agent reads the order queue and texts a customer the moment their coat is ready, something the owner would otherwise have to remember at 9pm between closing up and getting home.
Done carelessly, the same connection lets that agent process a refund on its own, waive a late fee because a message sounded annoyed, or quietly discard an order nobody collected. None of those are hypothetical mistakes. They're the ordinary kind an agent makes when it's technically allowed to do something a person should have decided. The fix has the same shape as the one Google chose for the front door: let the agent read and notify freely, and put a real person between it and anything that costs money, damages trust, or can't be undone with an apology.
Why "more connected" isn't the same as "more trustworthy"
MCP is plumbing. It gives an agent reach, not judgment. A tool that suddenly advertises "AI agent integration" hasn't necessarily gotten better at deciding what's wise: it's gotten better at acting on what it decides. Reach without judgment is a strictly worse combination than either alone: it fails faster and costs more when it goes wrong.
Google seems to know this. Its own announcement doesn't promise the integration will behave predictably. It admits results can vary depending on the agent connected and asks users to check carefully what access they're granting. That's an honest thing for a company to say about its own product, and it's the same honesty worth demanding from anything that offers to connect an AI agent to your invoicing, your bookings, or your customer messages.
The one question worth asking
Next time a tool you use offers to "connect your AI assistant," don't start with what it can do. Start with what it's not allowed to do by default, and whether you're the one who gets to set that list. If a vendor can answer that clearly, they've thought about this properly. If they can't, that tells you something too.
This is the same test behind What is an AI "agent", actually?: an agent is defined by what it's allowed to do without asking, not by how impressive its answers sound. Google just drew that line at your front door. Draw yours before someone else's software draws it for you.