100 Tools Down to 3: Prime Video's MCP Progressive Discovery

AI agents often suffer from performance degradation and hallucinations when provided with excessive tools, a phenomenon mirroring the paradox of choice. To mitigate this context bloat, Prime Video engineers implemented a progressive tool discovery mechanism within the Model Context Protocol (MCP). By utilizing a "find tools" function, agents dynamically load only the tools relevant to their current problem category, effectively swapping toolsets mid-session. This approach maintains a lean context window, significantly improving agent accuracy and reliability. The implementation relies on tracking agent sessions and mapping them to specific problem spaces, allowing for modular, scalable tool management. While this method introduces slight latency due to repeated tool list refreshes, it provides a robust solution for managing complex, cross-cutting infrastructure in enterprise environments where maintaining a clean, performant agent workspace is critical for operational efficiency.