(Should you) Drop everything and build AI Agents
I’ve been thinking a lot about how far we can go with AI Agents—specifically, whether to start building now or wait for the tech to “mature.”
Here’s my current thesis: starting today might be way more important than we think.
The Technology
Gary Marcus and others have pointed out that there might not be lasting moat in AI technology itself. Models get commoditized, APIs become interchangeable, and whatever breakthrough happens today gets replicated tomorrow.
Fair point.
The Work
For us mere mortals without billions of dollars to pour into R&D, the work in AI isn’t about doing years of in-depth research developing the foundation LLM models.
But here’s what actually takes time even when we get the model handed to us by the tech giants:
Building and deploying agents - Sure, writing some code, the infrastructure work, monitoring and maintaining the agents takes its time. However, This is cloud software development that we do know how to do. Equipped with modern coding AIs like Claude Code, this only becomes a real challenge at scale.
But more crucially…
Learning how to utilise agents in real organization - This is like learning to build and run an organization. We need to figure out:
- Which agents you actually need
- How to instruct agents effectively
- How agents should interact with humans and how we should interact with them
- How agents should interact with other agents
- What use cases actually deliver value
And here’s the kicker: we humans need days and weeks to think through these problems, both consciously and subconsciously. You can’t rush organizational learning. You can’t skip the messy iterations. You simply can’t get good at this overnight.
The Data
But there’s something else. Something potentially bigger.
With long-term memory storage, the real moat might end up being in the memories the agents produce.
Think about it: agents get better the more they’re used. More conversations = more memories = more learnings = better performance. And unlike humans, we can permanently store every single word of every conversation for just the cost of storage space.
Even better? We can improve how we distill memories later and literally re-run the entire memory-building process with a better strategy.
But we need that raw conversational data first. The sooner we start collecting it, the better.
Compounding Benefits
This is where it gets interesting (and maybe a bit scary).
Operating with agents might work like compounding interest in investing. Slow at first, but over time the improvements speed up. Each memory makes the agent better which helps the agent be more valuable, which helps the agent to be used more, which makes more new memories. Each optimization builds on the last.
If this is true, the cost of delay is massive.
The optimistic view: Start a little earlier than someone else, and you might build an ever-growing, unfair advantage.
The chilling view: Wait too long, fall behind, and you might never catch up to those who started early.
My Take
Do I know this will play out this way?
No. I really don’t.
But I think it’s extremely interesting (and potentially important) to see how things turn out and be part of playing it out.