Deferred Context Engine

Droid's Deferred Context Engine optimizes the handling of internal tools, MCP tools, skills, and plugins by loading only the necessary schemas and instructions when required, significantly reducing input context and improving efficiency. This approach, which has already been implemented in production, has demonstrated an average reduction of input tokens by 15.1% and even greater savings in sessions with extensive tool catalogs. By using progressive disclosure, Droid maintains a compact capability index and only fully loads schemas when a task demands it, allowing for more efficient management of large tool catalogs without overwhelming the system's context capacity. This method mitigates issues such as attention dilution, tool-selection noise, and premature context compression, thus enhancing the overall performance and reducing latency for users in professional enterprise setups. The system keeps frequently used tools readily accessible while deferring less common capabilities, ensuring that users can add tools and skills without the burden of carrying unused schemas in every task.