Agentic swam is the idea of having multiple AI agents doing multiple roles at once. Sounds like a convincing solution to our enterprise problem doesn’t it?

However it can’t be realistically applied at scale but our approach can. Why?

Full Autonomy

When you run an an agentic swarm to code, you’re essentially giving it full Autonomy. You could argue you can prompt it, but there’s no iteration involved with humans here. When it comes to concepts and big picture ideas, AI just doesn’t do the trick anymore. It only knows what it knows and I’d dare say it’s limited in it’s hallucinations.

Communication Issues

We have to communicate with each other to solve problems and same goes for agentic swarm. Can you imagine how much overhead this brings in when you have ten or hundreds of AI agents? It’s like trying to travel at the speed of light.

Information Overload

If we use the one AI approach on a certain part of the application, it can understand that easily. Due to AI tokenisation and memorisation limitation, swarm coding pushes this to the limit. Now for small projects it’s absolutely achievable, but if you’re looking at introducing this to a large already developed codebase – think again.

Error Application

Imagine your friend told you a fact, states it as a fact and you had no internet to confirm. You’d probably take it as fact. Or maybe you work your ideas around that fact when making decisions. This is what happens to AI. Mistakes can propagate and causes bigger issues down the road.

Coordination

When we code, we have tools at our control to help create better code, like Git, code reviews and project management. There is no such thing for AI right now, so these problems those tools solve won’t just disappear when we start going down the road of agentic swarming. The unpredictability of it all causes scaling issues.

Costs

Like I mentioned early, it’s like trying to approach the speed of light. Starting off might be great, but to scale, we’re looking at huge costs. Probably the cost of re-doing the entire codebase. Which many have tried when designing systems badly. Agentic swarm is good if you want it to do one thing well (the idea you give it) but for scalability, it’s a bit of a stretch.

Exciting Direction

Agentic swarm is an exciting direction indeed. However there are still real limitations as mentioned above, and probably not affordable for companies who need to introduce features regularly and iterate. At the crux of it, AI will still need orchestration and coordination from us. Abstracting these is great in theory but as for what the future holds, it looks unlikely to be taking over multiple developers with the power of living documentation, well designed systems and a single agent for each developer.

Check out the inspiration behind this post here