What's this about?
A safety rule and an episodic log compete for the same tokens in an agent's context. When the budget overflows, both get summarised at the same rate — but only the rule needs its exact wording to remain enforceable. The paper measures this across 20 production-style setups: the log survives summarisation fine, the rule quietly degrades into a paraphrase, and paraphrased rules stop being followed.
What is "compaction"?
Every model has a finite context window. When a long-running agent fills it, something has to go, so the harness summarises older content to make room — that is compaction. It works well for narrative history, where the gist is what matters. It works badly for anything whose *precise wording* is the point: a constraint, a policy, a credential-handling rule. "Never write to the production database" surviving as "be careful with the database" is a summary that is technically faithful and operationally useless.
Why does this matter to you?
- This is a concrete, testable failure mode in every long-running agent harness, including the ones you use daily.
- It reframes safety rules as things that must be *pinned*, not merely *included*.
- It explains a class of "the agent ignored my instructions" complaints that are usually blamed on the model.