# Headroom Context Compression Makes AI Agents Cheaper, Faster and More Accurate

This talk explains how making agent context more intelligent can improve cost, speed, and accuracy at the same time, with a focus on prefix caching and Headroom’s integration with Kong. The speaker argues that most tokens sent by agent harnesses are tool outputs and repeated history rather than user prompts, and that naive compression often breaks the prefix cache, increasing spend even when token counts fall. Headroom’s approach is to preserve the stable head of the conversation, compress only the changing delta, and use techniques such as lossless folding of repeated lines, structure-aware compression for JSON, code, logs, timestamps, and UUIDs, plus prose reduction with a non-generative model. Running locally as an open-source sidecar, Headroom integrates natively with Kong through before-and-after model APIs, adds about 18 milliseconds of overhead per turn, and is presented as delivering roughly 18 to 20 percent cost savings in demos while often improving accuracy by reducing bloated context.