Consumer AI Agents Are Already Working
Last week I asked an AI to plan a weekend trip for my family. Hotels, restaurants, kid-friendly activities, the works. I didn’t tell it where I live. I didn’t tell it my wife’s name. I definitely didn’t tell it she hates onions.
It figured all of that out anyway.
And when it delivered the itinerary, it had correctly scheduled nap time for 1:30 PM. For a one-year-old it’s never met.
This isn’t some VC pitch deck. It’s Google Spark, an AI agent that launched about a month ago and is now rolling out to Gemini’s $99/month Ultra subscribers. And I’ve been watching the reviews roll in from tech journalists who are, let’s be honest, professionally obligated to be skeptical about anything Google ships. They’re not being skeptical.
“I really said: ‘Wow, that’s actually nuts,’” wrote one Verge reviewer after Spark found his wife’s email, pulled grocery spending data from a budget spreadsheet that didn’t have “budget” in the filename, averaged the monthly totals including incomplete data from a month that wasn’t even over yet, and dropped it all into a draft email — complete with a personal sign-off that the couple apparently uses just for each other.
That’s not a carefully staged demo. That’s a Tuesday afternoon test that actually worked.
The Gap Nobody’s Talking About
We’ve been drowning in negative numbers about AI agents. 79% of enterprises say they’ve adopted them. Only 11% actually run them in production. Sixty-eight percent of the value just… evaporates somewhere between the pilot and the deployment. I wrote about this recently. The numbers are real and they’re bad.
But here’s what I keep thinking about: while everyone’s obsessing over the enterprise adoption gap, consumer agents quietly started working. Like, actually working. Not demo-working. Not “we’ll fix the hallucinations in the next release” working. Working right now, in production, for paying users.
Google Spark isn’t alone in this. It just happens to be the most visible example because Google threw a keynote at it and the tech press actually tested it instead of just rewording the press release.
Spark now runs on macOS, web, and mobile. It can sort your Downloads folder, create budget spreadsheets from invoices, connect to Canva and Instacart and OpenTable and actually use them. It tracks topics in real time — stock prices, sports scores, news — and pings you when something happens. Coming soon: you’ll be able to assign it tasks from your phone and it’ll execute them on your Mac while you’re away.
Is it perfect? God no. It couldn’t book an Airbnb because Airbnb’s authentication blocked it. It occasionally creates things that don’t exist, like a shared sign-up sheet for a block party it helpfully referenced in an email draft. The edges are still rough.
But the core loop — agent receives task, agent navigates real apps and real data, agent produces useful output — that loop is working. For real people. Right now.
The Quiet Revolution in Developer Tools
And then there’s the developer side, which is arguably further along.
Claude Code, Anthropic’s coding agent, has been out for months and the developer sentiment is… honestly pretty wild. People who were skeptical about AI coding tools six months ago are posting screenshots of agents refactoring entire modules while they go make coffee. SWE-bench scores keep climbing — Claude Mythos 5 hit 95.5% on the verified benchmark, which is the kind of number that makes “AI will assist developers” sound less like a prediction and more like a description of what’s already happening.
But the thing that actually excites me more than the benchmark numbers? The open source stuff.
Take OpenClaw. It’s a personal AI assistant you run on your own devices. Not in a cloud. Not behind an API paywall. On your hardware. It connects to 25+ messaging platforms — WhatsApp, Telegram, Discord, Slack, Signal, iMessage, you name it. You give it access to the tools you want it to have. It runs locally. It’s your assistant, not Google’s, not OpenAI’s.
A project like this shipping and gaining traction tells you something important: the agent paradigm isn’t just a big-company play. Regular developers are building and running agents that do real work — manage their inboxes, organize their files, keep them on top of calendar chaos — and they’re doing it with open source tools that didn’t exist 18 months ago.
Why This Matters More Than Enterprise Adoption
Look, I’m not saying enterprise adoption doesn’t matter. It obviously matters — it’s where the money is, it’s where the scale is, it’s where agents might eventually do the most good (or the most damage, depending on how governance shakes out).
But consumer and developer agents matter for a different reason. They’re proving out the paradigm in a way enterprise pilots never do.
An enterprise pilot has a project manager, a vendor relationship, a steering committee, a six-month timeline, and approximately 47 meetings about governance before anyone writes a line of code. It’s an artificial environment. You learn things about process and integration but almost nothing about whether the technology actually fits into someone’s life.
A consumer agent? The feedback is immediate and brutal. Either it books the right restaurant or it doesn’t. Either it finds the right spreadsheet or it creates a confusing mess. Either people keep using it after week two or they cancel the $99/month subscription.
And right now — mid-2026 — people are keeping the subscription.
The reviews of Spark are genuinely positive in a way that AI product reviews almost never are. Not “impressive for what it is” positive. Not “imagine what this could become” positive. Just… “I asked it to do a thing and it did the thing, sometimes in ways that surprised me.” That’s new.
The Invisible Infrastructure
Something else is happening underneath all of this that deserves attention. The cost of running these agents is collapsing.
Inference costs dropped roughly 80% from 2025 to 2026. Eighty percent. In one year. That’s the kind of cost curve that turns “technically possible but economically absurd” into “cheap enough to run in the background all day.” It’s why Google can offer Spark as part of a $99/month plan instead of a $999/month plan. It’s why open source agents can run on consumer hardware.
When the cost of a thing drops by 80% in a year, the set of things people build with it expands dramatically. That’s not a prediction. That’s how technology has always worked. Compute got cheaper, we got personal computers. Bandwidth got cheaper, we got streaming video. Inference is getting cheaper, and we’re getting agents that check on your stocks while you’re in a meeting and ping you when the threshold hits.
So Where Does This Go?
I don’t know. Seriously, nobody does.
The enterprise adoption stats are still grim. 97% of companies expect a major AI security incident. Only 6% have adjusted their security budgets. The governance gap is real and it’s not closing fast.
But the consumer story is different. It’s moving faster, it’s more concrete, and it’s less dependent on organizational change management — which, let’s be real, is where most enterprise tech goes to die. A consumer agent doesn’t need a steering committee. It just needs to be useful enough that people keep paying for it.
And based on what I’m seeing right now, they are.
Microsoft is building an entire operating system — Project Solara — designed specifically for agent-powered devices. It’s built on Android, not Windows, because the constraints are different when your primary interface is an agent instead of an app grid. AccuWeather, Best Buy, CVS, and Target are already planning hardware pilots. That’s not happening because agents are failing.
The 68% enterprise gap is real and I stand by every word I wrote about it. But if you’re only looking at enterprise adoption numbers to figure out where agents are going, you’re looking in the wrong place. The action right now is in the tools people are actually using — the coding agents shipping code, the personal assistants managing inboxes, the open source projects that put agent infrastructure in the hands of individual developers.
Consumer agents are already working. The 11% that made it to production? A lot of them are sitting in people’s pockets right now.