- How much context can deterministic reduction remove before correctness drops?
- When does semantic editing outperform patch/text editing?
- Which mutation metrics best predict agent performance?
- When should a supervisor delegate?
- How small can child context become?
- When does consultation improve recovery?
- When is provider fallback better than retry?
- How should price influence routing?
- How should measured historical quality influence routing?
- What authorization primitive fits autonomous workers?
- How should concurrent agents claim semantic code regions?
- How much does independent review improve correctness?
- Does destroying the source executor before verification reduce correlated errors?
- How much inference can event-driven waiting eliminate?
- How should discovery budgets scale with objective complexity?
- What durable state must actually be model-readable?
- When is deterministic software better than model cognition?
- When does moving behavior out of the model reduce useful adaptability?
- What is the minimum sufficient decision-evidence envelope?
- What mechanical operations are agent frameworks currently wasting inference on?