Curating Always-Loaded Context for LLM Agents: A Capacitated Assortment Model with Censored Feedback

Context curation for LLM agents is modeled as a capacitated assortment problem where instructions compete for a finite attention budget. The research proves an upper bound on optimal file size and demonstrates that naive appending of instructions can yield significantly lower net value than selecting an optimal subset.