The new generation of models—Gemini, LLaMA, etc.—have fantastic context windows.
They hold more complex information, for longer. That’s a big deal, right?
But... how good is good?
I’ve been testing the limits in real world rich context coding scenarios, and I don’t think this “high-context, high-recall” generation of multi-million token production ready AI's are fully baked yet.
Push them too hard and you’ll start to see weird behavior:
confusion, fragmentation, emotional meltdowns, random language switches, task-switching midstream.
Context overload is still very real.
It’s exciting stuff, but we’re not at the point where you should be building massive, fully functional monoliths using a single model, agent, or process.
You will hit complexity walls—especially on large codebases.
My advice?
- 🔹 Limit the context window to around 300 pairs
- 🔹 Use agentic memory and long-running processes when possible
- 🔹 Build around the “fits-in-your-head” principle
Because if you can’t hold a superhuman amount of working code in your head. Don't expect an AI to be able to do this reliably. At least... for now.
That’s why microservices matter.
- 🔹They reduce cyclomatic complexity.
- 🔹They isolate scope.
- 🔹And they soften the blow when your helpers get confused.
When there’s a will, there’s a way.
And when there’s a way, there’s a trick to it.
👉 Think in terms of components.