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

TEXT START: At the start of every session, LLM agents load a fixed context file, such as $\texttt{this http URL}$.

The Dissection

This paper turns persistent agent context from prompt folklore into an operations-research problem. Instructions consume scarce attention capacity, impose recurring session costs, and can interfere with one another. The paper formalizes selection, pruning, evidence thresholds, censored feedback, and limited reviewer capacity.

Its real function is deployment maintenance: preventing an agent’s institutional memory from becoming a junk drawer. The key finding is operationally real—appending every apparently useful rule can make the system worse, while deleting ignored rules can destroy latent value. This is a control layer for AI production systems, not an argument about human economic survival.

The Core Fallacy

The central DT failure is scale blindness. The paper optimizes the machine’s instruction surface while treating the surrounding labor system as irrelevant. Better context curation may improve agent reliability and reduce cost; it does nothing to preserve human productive participation. It can accelerate its collapse by making cognitive automation more capable and more economical.

At the local level, the model also compresses a strategic governance problem into a bounded assortment problem. It assumes that token capacity, instruction value, compliance, and feedback can be represented cleanly enough for optimization. In real deployments, objectives are contested, instructions can be adversarial, models adapt, and whoever controls the curation loop controls part of the agent’s effective power.

Hidden Assumptions

• Always-loaded context remains a major architecture rather than being displaced by retrieval, training, tools, or external memory.
• Attention capacity can be treated as a reasonably stable scalar token budget.
• Candidate instructions have measurable, sufficiently stationary values and manageable interactions.
• Session outcomes can be attributed to particular instructions despite model drift, confounding, and rare failures.
• An ignored instruction is not merely dormant, conditional, or protecting against an unobserved failure.
• The owner has a coherent objective; conflicts between users, operators, employers, and model providers are outside the model.
• Human reviewers remain available and their labor remains worth purchasing.
• The objective being optimized is operational value, not social distribution, bargaining power, or employment.

The abstract supports the paper’s local claims but does not expose enough of the proofs or assumptions to establish that its bounds survive broader architectures or strategic behavior.

Social Function

Classification: partial truth and transition management.

This is not copium. It does not claim that human labor remains indispensable. It makes AI deployment more governable by converting context failure into a measurable management problem. That shifts attention from displacement to optimization: from who is being removed from the production circuit to how efficiently the replacement system can follow rules.

The Verdict

A technically serious maintenance manual for the cognitive machinery replacing the post-WWII wage circuit. It is downstream of P1, indifferent to P2, and likely accelerates P3. The paper does not save human participation; it improves the allocation of machine attention, making the Sovereign’s capital more reliable while reducing the need for human cognitive labor. Under the Discontinuity Thesis, its importance is not salvation but confirmation: the frontier has moved from whether agents can work to how their scarce attention should be rationed.