Prompt Compression for Long-Context Agents 2026: LLMLingua-2 & Provence
Prompt compression is now its own layer in the agent stack: how LLMLingua-2 and Provence cut long-context token spend in 2026 without wrecking retrieval...
Chapter 1: Why Prompt Compression Became an Engineering Layer in 2026 For most of the LLM era, "prompt engineering" meant writing better instructions. In 2026, the harder problem is the opposite: deciding what not to send. Agents now run for hundreds of turns, pull from vector stores, scrape web pages, read repositories, and accumulate tool outputs that dwarf anything a human would type. The bottleneck moved from authoring tokens to selecting them. That selection step has hardened into its own layer of the stack, and the name it goes by is prompt compression. This is not a niche optimization. If...