Build a context compaction gate for AI agents: classify each tool output as keep, summarize, or discard before the next LLM turn, cutting token bloat and noise.