Engineering Agent Harnesses with Google ADK
Production Patterns for State, Governance, Multi-Agent Orchestration, and Observability
Angela Nguyen
Agents are easy to demo and hard to ship.
Every AI team hits the same wall: the agent nails the demo in a Jupyter notebook, then falls apart in production. Sessions lose context. Tool calls fail silently. There are no guardrails, no traces, no evals, no rollback. Security review says no. Your platform can't deploy it.
Moving an agent from demo to deployment requires more than a clever system prompt. It requires a harness.