Why AI Agents May Fail by 2025

Host

Welcome back to Tech Talks Today! I'm your host, Alex, and today we’re diving into a hot topic: AI agents. There’s a lot of buzz around them, especially with predictions that 2025 will be their breakthrough year. But our guest today, Utkarsh Kanwat, has built over a dozen of these systems and has some surprising insights. Utkarsh, welcome!

Expert

Thanks for having me, Alex! I’m excited to share my perspective.

Host

So, Utkarsh, you’ve been in the trenches building AI agents for various applications. What’s your take on the hype surrounding AI agents for 2025?

Expert

Well, I think a lot of what’s being said is overly optimistic. While I’ve built systems that deliver real value, the current excitement overlooks some critical mathematical realities.

Host

Interesting! Can you break that down for us? What are the key points that make you skeptical?

Expert

Sure! First, there's the issue of error compounding in multi-step workflows. If you have an AI agent performing a task, and each step has, say, a 95% reliability rate, the overall success rate drops significantly as you add more steps.

Host

Right, so if I’m understanding this correctly, 95% reliability sounds good in isolation, but when you're chaining tasks together…

Expert

Exactly! A 95% success rate over 20 steps only nets a 36% success overall. That’s far below what’s needed for production systems.

Host

So, it’s not just about how good the AI is at one individual task?

Expert

Exactly! It’s a mathematical reality. Even if you could miraculously achieve 99% per step, you still wouldn't meet the reliability required for production.

Host

Wow, that’s a real eye-opener. What about the costs associated with these systems?

Expert

Another crucial point! The costs can skyrocket due to context windows in language models. Each time an agent needs to keep track of a longer conversation or context, the computational cost increases exponentially.

Host

So, the longer the conversation, the more expensive it gets? Can you give us an example?

Expert

Sure! Imagine a customer service agent that needs to remember past interactions to assist a customer. As the conversation goes on, maintaining that context becomes prohibitively expensive.

Host

That sounds like a big hurdle for companies looking to implement these agents!

Expert

Absolutely. It’s a challenge that needs to be addressed. Moreover, successful agents I’ve designed rely on bounded contexts and human checkpoints, rather than trying to automate everything.

Host

So, it’s about finding a balance between automation and human oversight?

Expert

Precisely! Building systems that can operate effectively within our current limitations is crucial. Autonomous workflows are just not feasible at scale right now.

Host

It sounds like you’re advocating for a more measured approach to AI agents.

Expert

Exactly! Focusing on what works and acknowledging the limitations is key. The future of AI agents is bright, but it’s important to ground ourselves in reality.

Host

Thank you, Utkarsh! This has been an enlightening conversation, and I’m sure our listeners appreciate your insights.

Expert

Thanks for having me, Alex! I hope it helps people make informed decisions in the AI landscape.

Host

Absolutely! And for our listeners, stay tuned for more discussions on the future of technology. Until next time!