I’ve Been Using AI Agents Every Day. Here’s What the Hype Isn’t Telling You.
Last week, I was building a small internal tool for our office — something simple, a web scraper that checks a government portal daily and sends alerts when new circulars are posted. A year ago, this would have been a two-hour task: write the scraper, set up cron, configure email alerts, test, push. But I decided to try the new hotness. I fired up an AI coding agent, described what I wanted in natural language, and sat back to watch the magic.
Twenty minutes later, I had code that almost worked. The scraper ran. The cron was set up. But the email alerts went to the wrong address, the error handling was non-existent, and one of the libraries it used had a known vulnerability that any basic scan would catch. I spent another 40 minutes debugging and rewriting.
Don’t get me wrong — the agent saved me maybe 30 minutes total. That’s real productivity. But it wasn’t the revolution I’ve been reading about. And as I was fixing that scraper, three headlines crossed my desk that made me pause. All from the same week. All telling the same story, just from very different angles.
Zuckerberg’s Quiet Confession
On July 2, Mark Zuckerberg held an internal Meta town hall. The message wasn’t what you’d expect from a company spending up to $145 billion on AI infrastructure this year. According to TechCrunch’s Lucas Ropek, Zuck told staff that AI agent development had not “accelerated in the way” executives had hoped. Meta had already laid off 8,000 people and reassigned 7,000 more to an “Agent Transformation” division. The expected upside? It hadn’t “come to fruition yet.”
Let that sink in. One of the world’s most aggressive AI investors, with a budget larger than most countries’ GDP, is admitting that the AI agent revolution is moving slower than planned. He’s betting it’ll click in the next three to six months. But the restructuring — the layoffs, the reassignments, the massive bet — hasn’t paid off yet. And this isn’t some startup burning through seed funding. This is Meta. The company with billions of users, infinite compute, and arguably the best AI research lab outside of Google DeepMind.
If they’re struggling to make agents work at scale, what does that say about the hundreds of startups promising “autonomous AI agents for your business” right now?
I’ve written before about the security wake-up calls around AI coding agents with DuneSlide, and about how to build your first AI agent. But Zuck’s admission cuts deeper — it’s not about security or individual capability. It’s about pace. The industry promised autonomous agents by the end of 2025. Then mid-2026. Now we’re hearing “three to six more months” from the guy spending $145 billion a year. That timeline creep tells you everything about the gap between a demo and a production-ready system.
The Sandwich Shop That Proves We’ve Lost the Plot
Then there’s the other headline from the same day. Julie Bort at TechCrunch looked at Jersey Mike’s IPO filing — a submarine sandwich chain — and found that the word “AI” or “Artificial Intelligence” appears 22 times in its S-1 document. Twenty-two times. In a document about selling cold cuts on bread.
The company uses AI for… honestly, it’s not even clear from the filing. The mentions are boilerplate risk factors. “We are beginning to use AI Technologies in our business,” one reads. Like a disclaimer. Like they’re covering themselves because investors expect to hear about AI, even from a brand whose public face is Danny DeVito.
To put this in perspective: “Weather” appears 5 times. “Lightning” appears 0 times. A Jersey Mike’s location in Texas was actually struck by lightning in 2021. You’d think that’s a more realistic risk to disclose than an AI disaster for a sandwich franchise. But the market demands AI buzzwords, so AI buzzwords you shall have.
This is what a hype cycle looks like when it’s past peak. The signal has decayed into noise that even a sub shop feels compelled to broadcast. It reminds me of the blockchain craze in 2021 — remember when every company suddenly needed a blockchain strategy? Or the metaverse hype before that? The pattern is always the same: genuine technological breakthrough, massive investment, unrealistic promises, then a painful correction.
The Hidden Cost Nobody’s Talking About
And then there’s the third piece — the one nobody in Silicon Valley wants to discuss at parties. Tim DeChant, TechCrunch’s senior climate reporter, dug into the latest sustainability reports from Google and Amazon. The numbers are sobering. Google’s total carbon emissions are up 25% year-over-year. Amazon’s are up 16%. Google’s indirect supply chain emissions have doubled since its 2019 baseline.
The culprit? The data center buildout for AI. Amazon added more than 1.2 gigawatts of data center capacity in Q4 2025 alone — that’s a single quarter adding more capacity than most countries’ total. Google has started investing in natural gas plants — actual fossil fuel power plants — to keep its AI servers running. The renewable energy credits strategy is breaking down because solar and battery storage simply aren’t scaling fast enough to keep up with AI’s relentless appetite.
I find this personally frustrating. A few weeks ago, I covered how to block AI crawlers from scraping your website, partly because I’m tired of seeing my site’s content fed into models without consent. But the environmental angle is worse. We’re literally burning natural gas so that a search engine can generate an AI overview that most people don’t even want. Something about that math doesn’t add up.
My Take, For What It’s Worth
Let me be clear: I’m not an AI skeptic. I use these tools every day — Claude, various coding assistants, local models with Ollama (which I’ve written a guide on). They help me write better code faster. They help me brainstorm architectures I wouldn’t have considered. I genuinely believe AI is transforming software development in ways we’re only beginning to understand.
But here’s what I also believe: we’re in a massive correction period. The industry overpromised what agents could do in 2025. The gap between “AI that can write a to-do list app” and “AI that can autonomously manage your production infrastructure” is not a linear step — it’s a chasm. And companies that bet their entire headcount strategy on that chasm being bridgeable in 18 months are now facing the reality of physics, or at least of software engineering.
The Jersey Mike’s thing is funny until you realize it’s not just them. Every SaaS pitch deck now has an AI slide. Every startup that used to say “we’re Uber for X” now says “we’re AI-powered.” The signal-to-noise ratio has inverted, and retail investors and customers alike are going to get burned by products that have “AI” slapped on but no real intelligence underneath. I’ve sat through vendor demos where the “AI feature” turned out to be a simple if-else rule engine with a pretty UI.
And the environmental cost — that’s the part that keeps me up at night. I’ve been thinking a lot about how AI releases became political decisions. But they’ve also become environmental ones. Every time I spin up a local model on Ollama instead of calling an API, I’m making a tiny choice about efficiency. Most companies aren’t making that choice. They’re scaling first, asking questions later, and the planet is picking up the tab.
Where That Leaves Us
I think the honest take is this: AI agents will get there. The technology is real, the trajectory is positive, and the money flowing into the space means the talent and compute to solve these problems exist. But the timeline has been fiction from the start. Zuck knows it. The sandwich shop knows it. The sustainability reports are screaming it.
The next 12 months will separate the companies that build genuinely useful AI from the ones that just say “AI” on their marketing page. That’s a healthy correction. It’s just happening slower, and more painfully, than anyone wanted to admit.
As for me? I’ll keep using the tools that actually help — the coding assistants, the summarizers, the local models. But I’m going to be a lot more honest about what they can’t do yet. And I’m definitely not putting “AI” in my ICT division’s project descriptions just to sound cool. The sandwiches shouldn’t have to.