DruidAI’s presenters compare two banking agents through an intern analogy. An agent that escalates fewer conversations can look better on a containment dashboard while missing a deliberately designed handoff to a banker. The point is that escalation should be evaluated against the business purpose, not minimized automatically.
Their framework combines configured capabilities and runtime behavior under governance. Components receive scores for being absent, measurable or managed over time, with higher weights for governance and the agent mandate. The assessment covers boundaries, risk, compliance, observability, accuracy, access permissions, budget, shared handoff services, task context and explicit success criteria.
The presenters recommend repeatedly scoring live traffic against a human baseline to detect drift. They also propose measuring what work an agent removes and what that enables employees to do. Their industry containment figures come from DruidAI’s anonymized production data and illustrate different handoff designs rather than a universal ranking of agent quality.
Watch on YouTube
