“It’s the Architecture, Stupid” — Why Prompt Engineering Won’t Fix Agents

That’s actually a really interesting data point — and it might validate the thesis more than contradict it.

If the model’s “thinking” parameters are your bottleneck, it likely means the model is still being asked to reason through too much in a single pass. That’s exactly the problem the architecture is designed to solve: you break cognition into discrete skills with defined inputs/outputs, so each model call is a narrow, well-scoped execution — not open-ended reasoning.

With a smaller model like Gemma-4 E4B, the architecture becomes more important, not less. The model doesn’t need to “think” — it needs to execute a structured step. The cognitive load shifts from the model to the runtime.

Would be curious to know: are you running full skills with structured I/O, or using the model in a more traditional prompt-based flow on top of the framework? That distinction usually explains the bottleneck.