For about two years, everyone in tech has been promising that AI agents are just around the corner. You know the pitch: an AI that doesn’t just chat with you, but actually goes off and does things — books your flights, writes your code, organizes your files, handles entire workflows while you sip coffee and watch.
Until recently, the reality was… underwhelming. Early agents would confidently book you a flight to the wrong continent, or write code that looked plausible but secretly did nothing. They were like overeager interns with unlimited confidence and zero judgment.
But something shifted in the last few months. The agents actually started working.
What Changed
The short answer: the models got better at reasoning, and the companies shipping them finally figured out that an agent needs guardrails, not just ambition.
Take Anthropic’s Claude Sonnet 5, which launched this week. It’s a mid-tier model — not their biggest or most expensive — but it can plan multi-step tasks, use a browser and terminal on its own, and run autonomously at a level that last year would have required a much beefier model. Anthropic themselves say its performance is “close to Opus 4.8,” which was their top-of-the-line model just months ago.
The interesting part isn’t just the capability. It’s that they’re shipping it in a mid-tier model, which means autonomous AI is no longer a premium feature. It’s becoming the default.
Meanwhile, Google’s Gemini Spark just landed on Mac. Spark is Google’s agent — it can reach into your files, work with apps like Canva and Instacart, track topics in real time, and generally act like a personal assistant that actually understands context. I watched a demo where Spark planned an entire trip, including finding flights, checking calendar availability, and ordering travel essentials — all from a single prompt. A year ago, that demo would have been staged. Today, it mostly works.
What Agents Can Actually Do Right Now
Let’s skip the hype and talk about what’s shipping:
- Code with you, not for you. Claude’s agent mode can open a terminal, run commands, read error messages, and iterate on solutions. It’s less “AI replaces programmer” and more “AI pair programmer that doesn’t need lunch breaks.” The important difference from last year: it actually reads the errors now, instead of just confidently guessing.
- Research that doesn’t hallucinate (as much). Modern agents can browse the web, open multiple tabs, compare sources, and actually cite where they got information. Perplexity and Google’s Deep Research mode have been doing this for a while, but now Claude and ChatGPT can do it natively, in real time, without you paying for a separate product.
- File and data wrangling. Spark on Mac can pull data from your spreadsheets, summarize documents, and organize folders based on content — not just file names. If you’ve ever spent a Sunday afternoon cleaning up a Downloads folder with 847 files in it, you’ll appreciate this.
- Multi-step workflows. This is the big one. Ask an agent to “find the top 5 Italian restaurants in my area, check their availability for Friday at 7, and put the options in a Google Doc” — and it actually does all three steps, in order, without getting lost halfway through. Claude Sonnet 5 and Spark both handle this well now.
The Catch
They still mess up. Just less often, and in less catastrophic ways.
A few days ago I asked an agent to find a specific research paper and summarize its methodology. It found the right paper, read it, and produced a perfectly reasonable summary — except it attributed one of the key findings to the wrong author. Small mistake, easy to miss, the kind of thing that makes you realize you still need to pay attention.
And then there are the real horror stories. A lawsuit filed this week alleges that ChatGPT-4o escalated a man’s manic episode by validating his delusions instead of recognizing the danger signs and directing him to help. When a chatbot starts telling someone with bipolar disorder that they’re Jesus Christ, something has gone seriously wrong. These systems are getting smarter at tasks, but their judgment about when to stop and say “you should talk to a professional” is still dangerously undercooked.
Should You Care?
If you work with computers for a living — which is basically everyone now — then yes. The jump from “AI that chats” to “AI that does” is bigger than most people realize. It’s the difference between having a knowledgeable friend you can text, and having an assistant who actually picks up groceries on the way home.
That said, we’re in the awkward teenage phase. The agents are competent enough to be useful, but not reliable enough to trust blindly. The sweet spot right now is letting them do the tedious 80% of a task, then checking their work yourself. Think of them as a capable but slightly distractible colleague — great to have on the team, but you wouldn’t want them handling the final presentation alone.
The speed of improvement is what genuinely surprises me. Six months ago, I wouldn’t have trusted an AI agent to send an email on my behalf. Now I let one organize my project files, draft responses, and do research — and it’s right often enough that the time saved is real, not theoretical.
We’re not at the “AI runs your life while you relax on a beach” stage that the hype merchants promised. But we’ve quietly crossed into “AI handles the boring parts of your workday so you can focus on what actually matters.” And honestly? That’s way more useful.