AI agents aren't magic. They're distributed systems — with better marketing.

I built a simple multi-tool agent with MCP and Google ADK. Then I pushed it into production, and every "simple" step became a reliability problem: tools as independently deployed services, stdio meeting HTTP, side effects that worked fine on localhost failing in new ways.

Managed platforms have absorbed some of this since. The problems underneath haven't moved: protocol versions still drift, side effects still fail halfway, auth contexts still multiply with every tool.

This isn't a code walkthrough. It's a systems story — about what changed when the agent left localhost, and what's still true no matter what platform it's on.

If you're building agents meant to survive outside a notebook, this talk is about the parts no one shows in the demo.