Building agents that reach production systems with MCP

Production agents increasingly run in the cloud and need reliable ways to reach external systems. Three integration patterns exist: direct API calls (simple but creates M×N integration problems), CLIs (lightweight but limited to local environments), and MCP (Model Context Protocol), which provides a standardized, portable layer for cloud-hosted agents. The post covers best practices for building effective MCP servers: build remote servers for broad reach, group tools around intent rather than mirroring endpoints one-to-one, use code orchestration for large API surfaces, leverage rich semantics via MCP Apps and elicitation, and use standardized OAuth with Client ID metadata documents and vaults for credential management. On the client side, tool search and programmatic tool calling can reduce token usage by 85%+ and 37% respectively. Skills and MCP servers are complementary and can be bundled as plugins for domain-specialist agents. MCP SDKs recently surpassed 300 million downloads per month.