Your section on context rot is the absolute truth. I have been prototyping heavily with Claude Code lately, and learning to treat token attention like a strict budget was a hard but necessary lesson. More context does not equal better output. It usually just equals a highly confident but confused agent. Also, massive plus one to the AST-grep callout. Syntactically perfect but structurally doomed code is the silent killer in AI development right now. Really stellar writeup
Really glad you enjoyed Marissa! I 100% agree that context rot is one of the most important concepts on this list (so much so that I decided to take a deepdive on just that topic in my most recent article!)
Context rot maps surprisingly well to clinical workflows. An OR readiness agent with too much unfiltered chart and preference-card context could be confidently confused; the useful test is whether it keeps only the state that changes the next action, preserves uncertainty, and returns control when evidence conflicts.
Your section on context rot is the absolute truth. I have been prototyping heavily with Claude Code lately, and learning to treat token attention like a strict budget was a hard but necessary lesson. More context does not equal better output. It usually just equals a highly confident but confused agent. Also, massive plus one to the AST-grep callout. Syntactically perfect but structurally doomed code is the silent killer in AI development right now. Really stellar writeup
Really glad you enjoyed Marissa! I 100% agree that context rot is one of the most important concepts on this list (so much so that I decided to take a deepdive on just that topic in my most recent article!)
Context rot maps surprisingly well to clinical workflows. An OR readiness agent with too much unfiltered chart and preference-card context could be confidently confused; the useful test is whether it keeps only the state that changes the next action, preserves uncertainty, and returns control when evidence conflicts.