I Told Five AI Agents To Book Me a Flight. Only One Actually Pulled It Off.
Okay, so here's how this started. I was lying in bed around midnight, scrolling through my phone like an idiot instead of sleeping, when I saw yet another LinkedIn post about how "AI agents are replacing entire job functions." Booking travel, writing code, running customer support — apparently AI just... does it all now, no human needed.
And I got annoyed. Not because I don't believe AI has gotten good — I use it every day — but because nobody posting this stuff ever shows their work. It's all vibes and screenshots of a chat window saying "Sure! I'll handle that for you!" And then the post just... ends there. Did it actually work? Did the thing get done? Nobody says.
So I did something slightly ridiculous. I picked five AI tools that all claim to have "agent" capabilities — meaning they can go online, take actions, and get real things done, not just chat with you — and I gave every single one of them the exact same task, back to back, in one sitting:
"Book me the cheapest flight for next Friday morning, and actually complete the booking."
No extra hints. No hand-holding. No "oh wait, try this instead" when it got stuck. I just wanted to see, honestly, which one of these tools everyone's hyping up could actually finish what it started.
What happened over the next hour genuinely surprised me — and weirdly, it made me feel a little better about where AI actually is right now, instead of where the hype says it is.
Why This Actually Matters (Beyond Just Being a Fun Test)
Booking a flight sounds small, but it's actually a brutal test for AI. It's not a "give me an answer" task — it's a "go do something in the real world, deal with unpredictable websites, make a judgment call, and follow through" task. If an AI can't handle that, then all the talk about agents "replacing workers" is just noise.
I wasn't trying to embarrass anyone here. I genuinely wanted to know: has this stuff actually arrived, or are we still a year or two away from it being real?
The Lineup
I stuck to tools people have actually heard of, not niche stuff buried on some GitHub page:
- ChatGPT's agent mode
- Claude's browsing agent
- Gemini's built-in agent tools
- Perplexity's Comet assistant
- Manus, a dedicated autonomous-agent tool
Same phone, same instruction, one right after another. I made tea at the start. It was cold and forgotten by the end. That's how this went.
First Up: ChatGPT — Confident, Fast, Then Suddenly Shy
ChatGPT didn't waste time. It opened a browser, started comparing flights across a few sites, and within a couple of minutes handed me a shortlist with real prices, real times, the whole thing. I remember actually leaning forward a bit, thinking, "oh, this might just work."
And then — right at the payment page — it stopped. Politely asked me to enter my card details myself. Which, fair enough, nobody wants software swiping a credit card on its own. But it meant the job wasn't actually done. It found the flight. It didn't buy the flight. Big difference when the whole point was "book it."
Still, credit where it's due — watching it search felt less like a machine running a script and more like someone who actually knew what they were doing, just... didn't have the authority to hit "pay."
Then Claude — The One That Slowed Down to Ask a Real Question
Claude did something I wasn't expecting. Instead of racing ahead, it paused halfway through and asked me a question: did I want the absolute lowest price, even if it meant a brutal 6-hour layover, or something a little pricier that actually got me there at a reasonable hour?
My first reaction was mild irritation — I just wanted it to go do the thing. But sitting with it for a second, I realized... that's actually a good question. Nobody wants to save fifteen bucks and arrive exhausted at 2 a.m. before something important. It felt less like a tool blindly executing commands and more like it was actually thinking about the person on the other end of the request. Which, again, is a strange thing to feel about a chatbot, but here we are.
It also stalled at the payment step, same as ChatGPT. No booking. But somehow it felt less like a failure and more like it was being careful on purpose.
Gemini — Fast Start, Then Just... Stuck
Gemini plugged straight into flight search data and moved quickly — I'll give it that. For a second, I actually thought this one might be the winner. It got all the way to an airline's own booking page.
And then it hit one of those tiny "prove you're not a robot" checkboxes. And just... sat there. No error message, no crash, nothing dramatic. It just quietly stopped, like someone standing outside a locked door, patiently waiting for a key that was never coming.
There's something almost funny about that, honestly. All that processing power, stopped cold by a checkbox literally designed to stop exactly what it was trying to do.
Comet — Quick, But a Little Careless
Comet was, without question, the fastest of the bunch. Results back almost instantly. But speed came at a cost — it grabbed a few cheap-looking flights without really digging into the details, and one of its "deals" had a layover so long I'd have spent more time sitting in an airport than actually anywhere useful.
It never even attempted a booking. It just handed me a tidy list and stepped aside, like a very fast search engine wearing a nicer outfit. Useful for browsing. Not really built to finish the job.
And Then Manus Actually Did It
I'll be honest, this is the one I expected to disappoint me. Less buzz around it, less brand recognition, and I went in with fairly low expectations.
It searched, compared options, filled in traveler details without me typing a thing twice, and instead of freezing at the payment wall the way everything else did, it generated a secure payment link and sent it straight to me — one tap to confirm and pay.
Is that "fully autonomous"? Not exactly — a human still had to approve the final payment. But it was the only one of the five that carried the entire process through to an actual, real, confirmed ticket. No stalling. No mystery checkbox standoff. Just a clean process that ended with a confirmation email sitting in my inbox a few minutes later.
I actually said "wait, it worked?" out loud. Alone. At past midnight. To no one. That's a strange kind of moment — being genuinely surprised that software did the boring thing it promised to do.
So What Did I Actually Learn From All This?
Here's the honest version, no spin either way: agentic AI in 2026 is real, but it isn't magic. Every single one of these tools understood the task, searched intelligently, and made sensible calls along the way. A few years ago, none of that existed at all — so that part of the hype isn't fake.
But "fully autonomous" still comes with a human quietly hovering nearby, finger over the confirm button. And honestly? I don't think that's actually a bad thing. I don't want software silently spending my money without me knowing. What I want is for it to handle the exhausting, repetitive part — the comparing, the searching, the filling in the same details five separate times — so that all I have to do is glance at the final result and say yes or no.
That's the real story hiding under all the noise. It's not "AI replaces you." It's "AI clears out the boring 80%, and you handle the last 20% that actually needs a human brain."
Would I Trust an AI Agent to Book My Next Flight?
If you're picturing typing one sentence and waking up to a fully paid, confirmed booking with zero effort on your end — not yet, not reliably, not across the board. But if you're picturing something that saves you from opening six tabs at midnight, narrows things down with actual judgment, and leaves you with one final tap instead of forty minutes of comparing prices — that part is already here, and it's getting better fast.
I ended up actually using that booking, by the way. Real flight, real seat, real ticket in my inbox. Took five attempts to find the one tool that could see it all the way through — but it worked. And honestly, that's a pretty good place for this technology to be right now: not magic, not fake, just genuinely useful if you know which tool to reach for.
If you're short on time, here's what actually matters:
- AI agents today can genuinely search, compare, and reason through real tasks like flight booking — that part isn't hype.
- Most tools pause before spending your money. That's a safety choice, not a broken feature.
- Pick your tool based on what you actually want: speed (Comet), thoughtful questions (Claude), tight integration (Gemini), balance (ChatGPT), or something that actually finishes the job (Manus).
- We're not at full autonomy yet. Think capable assistant, not replacement for your own judgment.
Comments
Post a Comment