Asynchronous AI Agents and Human Behavior | Designing Agents

AI Agents aren’t just typical programs. If you start thinking of them as having similar capabilities to humans, they get simpler in theory.

I’m not saying AI agents have the thing that makes you and I human. They aren’t living. They’re just very fancy, extremely convoluted programs. What I am saying is: shift your perspective on how they operate.

When you start comparing AI agents to how humans operate and you write them to behave in similar patterns they become much easier to conceptualize and plan.

One night while working on an asynchronous system for an agent framework, I had an epiphany.

While working on the Kernl system, I was stuck in the traditional software mindset: How do I create triggers for when agents should come back and do tasks? Webhooks? Maybe, but users don’t want to write configs for every webhook their agent might need. There had to be a better way.

That’s when I landed on a sleep and wakeup system.

Say you want this workflow:
1. Email Bob “Hello”
2. Wait for Bob to respond
3. Email back “Goodbye”

How would a human do that?

You’d immediately send “Hello” because that part is instant. Then you’d check your email to see if Bob replied but you wouldn’t sit there refreshing your inbox recursively until he responds. You’d check, realize he hasn’t replied yet, and come back later. You’d check at intervals.

So why not let agents behave the same way?

The agent checks the inbox, sees: “Oh, Bob hasn’t replied yet. I should wait.” Then it sets an alarm to wake up in an hour and goes to sleep. An external system wakes it up, it checks again, and repeats until Bob responds.

The important part here is that agents aren’t truly “real-time” in the way people expect software to be now. There isn’t a global framework for instant notifications. Humans only get instant notifications because our phones act like that framework. Without smartphones, we’d handle this the same way: check, wait, check again.

If you plan agents around this idea, they become easier to design. And you can flip the usual complaint “it’s not real-time” into an advantage: reliability and uptime.

If you hire an employee to monitor an inbox, they’ll do it 8 hours a day. They get distracted. They won’t check consistently. An agent can check every 30 minutes, 24/7, forever, without drifting.

That’s the mental model shift: less like event-driven software, more like a person with a perfect routine.