How Dash uses context engineering for smarter AI

Dropbox Dash evolved from a traditional RAG-based enterprise search into an agentic AI system, requiring a new discipline called context engineering to manage what information models receive.

• •Too many tool definitions caused analysis paralysis, degrading decision quality; Dash replaced multiple retrieval APIs with a single unified search tool backed by a universal index.
• •A knowledge graph layered on the Dash index connects people, activity, and content across sources, enabling relevance-ranked filtering so models only see meaningful context.
• •Complex tools like query construction were offloaded to specialized sub-agents with focused prompts, freeing the main planning agent to focus on task execution.
• •Context rot—accumulated tool call overhead degrading accuracy on longer jobs—was a key motivation for limiting token-heavy MCP tool definitions.
• •The Dash MCP server exposes just one tool to external apps like Claude and Cursor, keeping descriptions lean to maximize usable context