A pilot running one coding agent against one fixed task under six escalating levels of project context. Every environment finished, the richest ones spent up to 36 percent fewer tokens, and caching served 88 to 96 percent of the input.