Post
The Smart Home Was the Beta Test for AI Agents
Every decade tech promises the same thing: 'your stuff will finally work together.' Smart home, IoT, now AI agents. The shape of the failure is identical — and the way out probably is too.
So, when's the last time your smart home worked? Like, all of it. Lights, thermostat, locks, speakers, the cute little plant sensor — all responding to the same voice command, no "the bridge needs an update," no "this device is offline," no app you have to keep around just to control one bulb you bought at IKEA in 2019.
Yeah. Same.
Now look at what's happening with AI agents.
Every demo shows an agent that books your travel, files your taxes, drafts your code, orders the right size of dog food, and politely apologizes when it makes a mistake. The diagram has tidy arrows. The tools are all "MCP-compatible." The agents "delegate to each other" via A2A. The future, we're told, is composable.
We've watched this exact movie before. All of us have. It was called the smart home, and before that it was called IoT, and before that it was called Web Services, and at every step the pitch was the same: your stuff will finally work together.
It mostly hasn't. And the agent wave is currently running through the same chapter list.
Three Eras, One Plot
Strip out the buzzwords and the eras look almost identical:
| Era | Roughly | The Promise | Whose Hub Wins | How It's Going |
|---|---|---|---|---|
| Smart Home | 2010–2018 | "Your devices will all work together" | Nest vs SmartThings vs HomeKit vs Alexa | Z-Wave (1999), Zigbee (2003), Thread (2014), Matter (Nov 2022) — four standards in to fix the first one |
| IoT | 2014–2020 | "Everything will be connected" | AWS IoT vs Azure IoT vs every consortium ever | Mirai (2016) DDoS'd half the internet using default IoT passwords |
| AI Agents | 2024–? | "Your software will work together via natural language" | OpenAI vs Anthropic vs Google vs LangChain | MCP launched Nov 25, 2024. A2A April 2025. Both are version 1.x. We're early. |
Same promise, different layer of the stack. Smart home connected light bulbs. IoT connected sensors. Agents connect software functions. The wires are abstract this time, but the wiring problem is the same one.
The Five Things That Always Happen
Every wave goes through the same beats. We've now seen them three times. These are the load-bearing patterns:
1. The early demos work because there are five devices. SmartThings was magic when you had one hub, one bulb, and one motion sensor. Mine still works with five things. It got weird at fifteen. By thirty, something is always offline. Today's "agent that books your travel" demo has three tools wired up. The version with thirty MCP servers behaves differently — and worse — than the demo. The Nx integration problem isn't solved by a protocol; it's exposed by the protocol.
2. Everyone wants to be the hub. Smart home: Apple, Google, Amazon, Samsung, Nest, Wink, Lowe's, Philips, IKEA — all of them sold you a box that is the hub. Agents now: every model provider wants its model to be the orchestrator. Every agent framework wants to be the planner. Every IDE wants to be where the agent lives. The hub war doesn't end with a winner. It ends with a protocol nobody loves.
3. The protocol arrives late and weak. Matter shipped in November 2022 — about a decade after smart home went mainstream. It still doesn't cover cameras or robot vacuums properly. MCP (Nov 2024) and A2A (Apr 2025) are right on schedule for "useful but missing the hard parts." The hard parts in agents are state, identity, billing, and trust — exactly the parts neither protocol solves yet.
4. Companies die. Your stuff dies with them. Lowe's bricked Iris in 2019 and gave people Visa gift cards for their now-paperweight devices. Wink kept asking for $5/month. Revolv was killed by Google. Every "smart home cemetery" article has a hundred entries. The agent equivalent: the startup whose MCP server is wired into your workflow today gets acquired, pivots, or quietly stops shipping. Your "agent that does X" stops doing X. You'll discover this when your taxes are due.
5. Security is an afterthought, then a catastrophe. The smart home wave gave us Mirai: a botnet built from 600,000 IoT devices using a list of 62 default passwords. On October 21, 2016, that botnet took down GitHub, Twitter, Netflix, Reddit, and Spotify by aiming at one DNS provider. Default credentials, no patching, no thought given to what happens at scale. The agent wave is busy giving LLMs database access, shell access, and money. We are speed-running this chapter. Prompt injection is the new default password.
What's Actually Different This Time
Agents won't fail the way Iris did. A few things genuinely change with LLMs in the loop, and they cut both ways.
The translator is now elastic. In the smart home era, two devices either spoke the same protocol or they didn't. With an LLM in between, "almost compatible" can sometimes work. That's a real upgrade. It's also why the failures are weirder: instead of "device offline," you get "agent confidently did the wrong thing in the right format."
The integration cost moved from labor to tokens. Building a Z-Wave-to-Zigbee bridge took a small team of firmware engineers. Building an MCP server is an afternoon. Lower bar means more integrations exist — and more of them are unmaintained.
Failure is now stochastic. In the smart home, your light bulb either turned on or it didn't. You knew. With agents, the same prompt can succeed nine times and quietly do the wrong thing on the tenth. That is genuinely a new failure mode, and it makes the "your stuff doesn't work" complaint harder to debug.
What the Smart Home Era Already Told Us
If you take the parallel seriously, a few things are pretty obvious:
- The winners will be the patient utilities, not the orchestrators. The smart home wave's actual winner wasn't Nest or SmartThings — it was the boring Zigbee bulb that still works ten years later because it didn't try to be the hub. In agents, the equivalent is going to be the unsexy MCP server that keeps shipping for five years.
- "Open" matters more in year three than year one. Z-Wave was proprietary. Zigbee was a mess. Thread + Matter took over because they're truly open. The agent wave's protocol layer will end up at the same place. Bet against anything that requires you to be on a single vendor.
- Pick a hub you can replace. Everyone whose smart home survived 2010–2025 chose hubs they could swap (Home Assistant, late-game SmartThings) over hubs they couldn't (Iris, Revolv). Same logic applies: don't wire your life into an agent platform that can't be migrated.
- Assume a security incident is coming. It always does. The cleanup is what separates the protocols that survive from the ones that don't.
Where This Lands
The agent wave isn't going to die. Neither did the smart home — most houses now have at least one connected device, and those devices mostly work. But the version we end up with isn't going to look like the demos.
Five years from now, "AI agents" is going to mean a few quiet, boring utilities that just work — probably in narrow domains like calendars, code review, and customer support. The "agent that runs your whole life" is going to be discussed at conferences the way the "fully connected smart home" still gets discussed: aspirationally, by people whose lights still need a power cycle on Tuesdays.
That's not a bad ending. It's just not the one in the keynote.
We've seen this movie before. We're allowed to enjoy it without believing the trailer.
Sources
- Mirai (malware) — Wikipedia — primary timeline of the 2016 IoT botnet, the Krebs and Dyn attacks, and the source code leak
- Inside the infamous Mirai IoT Botnet — Cloudflare — retrospective on what Mirai did and why it worked
- Introducing the Model Context Protocol — Anthropic — MCP launch announcement (Nov 25, 2024)
- Announcing the Agent2Agent Protocol — Google Developers Blog — A2A launch (April 2025)
- Lowe's is killing off and bricking its Iris smart home products — TechCrunch — case study of the smart home shutdown and brick problem