Trimming an agent's context below 25% raises failure odds 10.92x, and protocol-aware trimming beats recency or summarization by 5.24x

arXiv 2609.16461·high signalXBlueskyLinkedInCopy link

A comparison of five context-trimming strategies across retained-context levels and workflow-complexity classes finds conventional approaches (recency, relevance, summarization) hit about 60% token savings but drop task success to 66.6-77.3% and protocol adherence to 85.5-88.6%. Protocol-aware trimming lifts success to 92.2%, and adaptive budget guardrails reach 96.0% success, 96.3% protocol adherence and a 1.0% cascading-failure rate while still saving 56.0% of tokens. The threshold to hard-code: retained-context budgets at or below 25% carry 10.92-fold failure odds versus 50% or more, and the critical threshold rises with workflow complexity, so a fixed compaction percentage is the wrong control.