Last week we hosted a livestream that turned into one of those conversations where you forget you're recording because you're just genuinely fascinated by what someone's building. My buddy Dennis Oleksyuk, CTO and co-founder of AirCon, joined me to talk about AI agents, agentic coding, and why most of what you read online about building production AI systems is basically garbage.

Dennis and I go way back to our DataRobot days, where we built some genuinely cool stuff together (including a pitch predictor that may or may not have gotten smartwatches banned from MLB games, but that's a story for another time). These days, he's neck-deep in one of those industries you never think about but that literally makes the modern world function: air freight forwarding.

The Industry You Didn't Know Existed

Here's something wild. Every time you've flown on a passenger aircraft, there was freight in the belly of that plane. Not on some separate cargo flight. Your flight. And there's this entire invisible industry of freight forwarders who orchestrate getting pallets from, say, Indonesia to Sacramento by stitching together 12 to 15 different specialized companies, each handling one tiny piece of the logistics chain.

It's complicated work. It involves custody chains and TSA regulations and knowing which airline's planes have doors big enough for your shipment. And until very recently, it was all done by humans spending half their day writing quotes via email.

Enter the AI Agent

AirCon built an AI agent that reads incoming quote requests, figures out the optimal route, checks which airlines can actually handle the shipment dimensions, adds margins, and sends back a complete quote. Autonomously. Then, when the customer says "book it," another agent goes and actually books everything across all those different providers.

The kicker? It's handling about 80% of air freight quotes completely on its own. The freight forwarder wakes up in the morning and discovers they've already made money while they were sleeping.

Why Most Agentic Frameworks Are Useless

This is where Dennis dropped some truth bombs that every engineer building AI systems needs to hear. If someone tells you they're using LangGraph or another popular agentic framework in production, Dennis says he'd bet money they're not actually in production. And he's probably right.

The problem isn't that these frameworks are bad. It's that they're optimized for rapid prototyping, not production reliability. They're bloated with every possible approach people have tried, when in reality you only need three of those approaches to actually work. You end up learning special terminology and wrestling with framework constraints instead of just solving your problem.

Dennis's recommendation? Take the raw LLM API and orchestrate it yourself. Yes, it's more work upfront. But when you need to re-architect something (which you will, weekly), you're not fighting against framework assumptions.

The Dirty Truth About Context Windows

Here's something most people don't understand: those massive context windows that AI companies love to brag about? They're basically just RAG (Retrieval Augmented Generation) moved into the model. Marketing term, not magic.

When you load a million tokens into a context window, the model isn't actually reasoning over all million tokens. It's distilling them through a much smaller bottleneck. Sometimes it generalizes everything. Sometimes it picks pieces. But if you actually need the model to think through all that information? Good luck.

This is why someone can write a two-page prompt and get worse results than someone who writes two sentences. More information doesn't always help. Sometimes it just clogs the bottleneck.

The Corner Cases Will Kill You

The fascinating thing about building AI agents for real business processes is that accuracy isn't actually the problem anymore. Dennis pointed out that as of August 2024, frontier models basically don't make extraction errors if you prompt them correctly. They're essentially perfect at reading data.

The real problem is corner cases. Every business process has an infinite number of them. Humans handle corner cases by understanding first principles, by having a mental model of how the world works. LLMs can't do that reliably in business contexts. They can fake it for toy problems, but when you need high reliability and repeatability, they just don't have the ability to reason from first principles.

So you have two options: build recipes for every corner case you discover, or teach your agent to recognize corner cases and escalate them to humans. That's where the moat is. AirCon's competitive advantage isn't the AI, it's all those corner cases they've catalogued and solved.

What Actually Matters

Look, building AI agents that work in production comes down to understanding your domain well enough to handle the corner cases. Not the happy path. The weird stuff. The thing that happens once every hundred transactions that breaks everything if you don't account for it.

You can build something that works 80% of the time on a single weekend. Getting to 95% takes months of cataloging edge cases and figuring out why the model chokes. That's the gap between a cool demo and a system that makes money.

Here's what I took away: stop obsessing over which framework to use. Start obsessing over the business process you're trying to automate. The LLM is infrastructure. The hard part is understanding the domain deeply enough to know when the AI can handle something and when it needs to tap out and get a human involved.

Dennis has freight forwarders waking up to discover they made money while they slept. That doesn't happen because he's using the coolest framework. It happens because he spent months figuring out which airlines fly which planes between which cities and teaching his agent when to escalate. That's the unsexy work that actually matters.

Watch the recording here: