Senior Backend Engineer · Java, Go & AI (MCP) · Building production-grade systems
Senior Software Engineer with 14+ years building backend systems in Java (8 to 25) and Spring Boot, now also working with Go. I write about MCP, AI agents, cloud-native architecture, observability and taking systems to production. Based in Colombia.
Markdown supported · ⌘↵ to post
OneLifeVibe
The framing of MCP as just another driving adapter is the part that stuck with me. It makes the hexagonal approach feel practical instead of abstract, because adding a new entry point does not mean rethinking the domain. I also liked the point about tool descriptions being the real API, and the example of findOrder over queryDatabase. That is a small naming choice with a large effect on how reliably an agent calls the right tool. The ArchUnit guardrails and the observability section round it out well. How did the k6 load test shape the final tool design, if at all?