Google Developers Blog published "Evaluating AI agents: moving past the vibe check" on October 15, arguing that agents fail differently from ordinary software: one can get stuck in a retry loop and fire off multiple tool calls yet still return a polite, plausible-sounding final answer, quietly spiking token costs, exhausting downstream API quotas and triggering duplicate side effects. The authors recommend anchoring tests in critical user journeys rather than abstract benchmarks, starting by reading 5 to 20 execution traces by hand with a "red pen" and turning those notes into deterministic assertions and boolean pass/fail judge checks. They lay out a Day 0 to Day 2 rollout using the open-source agent-eval toolkit to capture OpenTelemetry traces on OpenInference conventions and route them to the Agent Evaluation service on the Gemini platform, with checks eventually gated in CI/CD.