Seven Blocks of Debt Hiding in Agent Infra

Port's CEO Zohar Einy mapped out seven categories of infrastructure debt that accumulate once AI agents move past the prototype stage - from stale context files and unscoped credentials to invisible agent sprawl across Claude Code, Cursor, n8n, and Zapier. The real problem isn't building agents; it's that teams wire up integrations independently, approval logic gets hard-coded per agent, and decision traces vanish after each run. By the time multiple teams are running agents in production, roughly half the engineering capacity goes toward infrastructure nobody planned for. Google flagged similar maintenance traps in ML systems back in 2015, but agent-specific issues like non-deterministic orchestration and fuzzy inter-agent contracts are a newer category entirely.