Making Agents Reliable Enough to Leave Alone

Drew Donaldson

Published on: 07/07/2026

Moving an AI agent from impressive demo to running unsupervised is a measurement problem, not a model one. Here is what eval harnesses, guardrails, and observability actually take.