How Dash Uses Context Engineering for Smarter AI

Dash started out as a typical RAG system for document search, but it has now evolved beyond simply finding content into an agentic AI that interprets, summarizes, and even takes action. The core challenge along the way was context engineering—making sure the model sees only the information and tools it actually needs.

As the number of tools grew, options increased, but decision-making actually slowed down and accuracy dropped. Approaches like MCP, which inject each tool's description and parameters as-is, quickly consumed the context window, and context rot even appeared in long tasks. To address this, Dash adopted the following three core principles.

• Limit the number of tool definitions.
• Filter for only highly relevant context.
• Separate complex tasks into dedicated agents.

The first strategy is not to over-proliferate search tools. Instead of exposing the search APIs of multiple external apps to the model individually, Dash consolidated them into a single consistent search tool based on a universal search index. This approach keeps the model from getting lost among multiple APIs, and also lets Dash offer search functionality in a more compact form to external MCP-compatible apps, such as via the Dash MCP server.

The second strategy is to not pass search results through as-is, but instead leave only pre-organized, highly relevant information. Dash bundles data from multiple sources into a single index and layers a knowledge graph on top of it that connects people, activities, and content. This way, context doesn't need to be reassembled at runtime, making things faster, and the model only sees meaningful information that has already been curated.

The third strategy is to pull out tools that are too complex into dedicated agents entirely. Search query construction in particular required a lot of explanation, since it involves interpreting user intent, matching it to index fields, and handling typos, synonyms, and implicit context. Dash separated search out from the main planner, designing it so the main agent focuses on planning and execution while a dedicated search agent handles only query generation.

The core message is clear: agentic AI is not a matter of feeding in more context, but of feeding in more appropriate context in a better form. Through tool consolidation, relevance filtering, and separating out specialized agents, Dash avoids scattering the model's attention and instead lets it spend more compute and reasoning on the tasks that truly matter.

Going forward, Dash plans to expand by refining even more context, including user and company profiles as well as short- and long-term memory. The authors believe that MCP itself is valid, but that scaling it up requires reducing tool sprawl and, where needed, even leveraging code-based tool generation.