I’ve tried Astra in a few different environments and have started tweaking my setup because of it. What I noticed is that the model performs worse in harnesses with a lot of scaffolding and large system prompts (I was using oh-my-pi). It takes a while to get any work done because it keeps burning through context and iterating on things that could have been one-shotted.
On the other hand, when I did a fresh install of Codex and tried Astra with no skills or instructions, it performed much better. Tasks were completed faster and with less input from me.
It used to be a flex to have a complicated setup that looked sophisticated, but now it seems to be getting in the way of model performance. As general intelligence grows, skills, large prompts, and verbose documentation will probably push models toward anti-patterns that were designed early on, when a better approach wasn’t available.
Integrating with an existing codebase still matters. But heavy documentation just becomes slop overtime that is not useful. This is most likely easier to fix locally, but for products where you don’t see the internal prompt its kinda like a hidden defect.
I’ll probably keep my setup minimal going forward. It shouldn’t be necessary to write an essay explaining my workflows and design concepts. I’ve found it easier to just list the design patterns and let the model figure out the intention behind them.