AI Agents Are The New Flying Cars
2025.19 - The reaction from TPMs I have talked to using the current crop of AI agents tools is mixed, some disappointing. Problem isn't the tool, it's the hype, expectations, and approach.
We’ve all heard it by now: “AI agents will change everything.” And maybe they will. Maybe for some areas of tech industry they already are. For many, that change is still a promise far way.
There’s this quiet disappointment echoing in teams experimenting with agents.
People build their first assistant, deploy it, and it doesn’t work as they expect. It hallucinated. It got stuck in a weird place. Throws “don’t have access to that data source”. You prompt engineer for the next several minutes to hours. You might get 80% of the way to the original goal. Eventually, some teams back off. “I guess we’re not ready for this.”
But here’s the thing: That first failure? It’s not a flaw in your AI journey. It is your AI journey.
Your First AI Agent Will Fail. It’s Ok. Keep going.
The more I experiment with AI Agents and Tools, the more I think about Flying Cars
I mean think about it, we’ve been building cars and planes for over a century. We have sent both to the moon. You’d think putting the two together would be “simple enough”. But flying cars aren’t just “cars with wings.” They’re entirely new systems with conflicting design needs, complex safety challenges, and high coordination costs. It’s not about what you build. It’s about how well it works together.
Same with AI agents.
We’ve had LLMs for a few years. We’ve had APIs, databases, task managers, and user interfaces for decades. But stitching them into a dependable autonomous system? That’s still hard. Requires design and systems thinking. Experimentation and actual problem to tackle.
And that’s where most teams give up, at the very moment they should lean in.
A Thought On Investing In AI Tools For Team Productivity
Look, the hype is super strong. If the recent Y Combinator class is any indication we will have more startups promising simple easy to use agentic tools. Maybe at some point we will get there. For now, treat AI agents not as products. They are programs that needs careful planning and investment and rapid iteration.
Here’s the mindset shift that TPMs need to bring to the table:
Don’t treat agents like one-off tools.
Treat them like multi-phase programs.
That means to seriously deploy agentic tools to help teams:
Starting with narrow, observable tasks; don’t go for system level workflows from the get go.
Tracking what fails and why because those failure points are your real design requirements.
Thinking in versioning: v0.1 isn’t bad, v0.n incrementally better.
Defining ownership and review/feedback loops on what is working and not.
Iterate, iterate, iterate.
Agents break because we approach them like static features or all encompassing SaaS tools ready to go out of the box. But they’re dynamic systems. They change as your workflows, data, and context change. They are effective on what they have access to and even then they need to be explicitly told what to do.
Which means: you’re not launching a product. You’re managing a living process. Even more, the tools and the technology that powers them are changing on a regular rapid basis.
💡 For TPMs: Start the program. Expect the failure. Learn out loud.
If your org is exploring AI agents, take the lead as a TPM:
Frame it as a program.
Run it like any other transformation: pilot, evaluate, iterate, scale.Redefine success.
Not “did it work flawlessly?” but “did we learn something actionable?”Create failure rituals.
Postmortems, logs, bug bashes. Normalize the learning curve.Start embarrassingly small.
A bot that summarizes Slack threads well is more useful than an over-engineered assistant that half-writes PRDs.Set the expectations clearly.
Don’t McKinsey or Klarna it. Set clear expectations what the is possible and bring everyone along on the journey.
Final Words
When these tools work, it’s like WOW! When it doesn’t work on the most obvious things, it’s like THAT IS SO UGHHHH.. DUMB! The more I experiment with these tools the more I keep telling myself the point isn’t to get it right but it’s to get it going.
Every mature system starts with immature experiments. I fully expect your first AI agent will fail. Your first initial experiments will result in disappointing and limitations.
That’s not a reason to stop. That’s your invitation to build something that actually works. Not a one-off assistant. But a real program with direction, resilience, care, patience and investment.
That’s where TPMs shine. This is why I say that if your company is not experimenting with any Gen AI Tools beyond “We use ChatGPT”, that is your cue to lead the way.
Until next time.
-Aadil
P.S. There is a great presentation by Andrej Karpathy that honestly is sublime and captures the mindset required to approach this new crop of technologies. Worth the 40mins of your time.
Don't Klarna it! I'm stealing that one... :D