ACE (Agentic Context Engineering): Delta-Update System Prompts for 10.6% Benchmark Gain

SoftmaxData Blog·high signalXBlueskyLinkedInCopy link

ICLR 2026 paper showing that treating system prompts as living, incrementally-updated documents yields 10.6% improvement on AppWorld benchmarks and enables smaller open-source models to match production closed-source systems. Technique: structure prompts as itemized collections with metadata, apply targeted delta updates when failures are observed rather than rewriting from scratch, implement reflection cycles that append failure-derived tips, and consolidate accumulated knowledge periodically. The key insight is that a prompt that learns from its own failure history outperforms a carefully hand-crafted static prompt.