Paper Models AI Agent Context Files As a Capacitated Assortment Problem

Researchers Zexuan Liu, Yuning Yang and Tiancheng Zhao posted a preprint on arXiv treating the context files that large language model agents load at the start of every session, files such as AGENTS.md, as a resource allocation problem rather than a static text block. Their framing is that each instruction kept in the file is loaded again in every later round of the session, so a file built by appending instruction after instruction keeps charging the same token cost over and over, and the paper reports that bloated files can degrade an agent’s performance as they grow.

The core contribution is a capacitated assortment model: instructions compete for a finite amount of attention capacity, adding one instruction never raises how well the agent follows the others already present, and each retained instruction carries a per-session setup cost on top of its token cost. The authors prove an upper bound on how large the optimal file can get no matter how many candidate instructions exist, and they show that appending every instruction with some standalone value can produce a worse outcome than deliberately choosing a smaller subset. A capped token budget, they find, also limits how badly things go wrong when the price of a token is underestimated.

The second half of the paper turns to how curators, human or automated, should learn what belongs in the file. Feedback on this kind of document is inherently incomplete: an instruction that is loaded generates visible signals of help or harm, while an instruction that was never added only produces feedback once its absence has already caused a problem. The authors show that deleting instructions an agent appears to be ignoring can end up removing instructions that were actually useful, and they set out how much evidence a curator should wait for before adding a new rule. They also bound how much a reviewer’s workflow suffers when only a limited number of edits can be inspected per period.

The numbers

• The paper proves a fixed upper bound on the optimal size of an always-loaded file, holding regardless of how many candidate instructions a curator has to choose from.
• Its experiments test the model against real context files rather than synthetic text, and report that irrelevant rules drawn from those files reduce how well a language model complies.
• The censored feedback result shows that an instruction’s absence produces a signal only after that absence has already caused harm, not before.
• A bound is also derived for a reviewer who can inspect only a capped number of edits in each period, rather than reviewing every change.