Search the Atlas

Search risks, controls, and glossary terms

AgenticGoal & Objective Control

Multi-dimensional objective specification with intent audits

Control objective

This control ensures that every metric-driven agent operates from an explicit, documented statement of what it is supposed to achieve and within what limits - its 'intent spec' - rather than chasing a raw optimisation target that it might satisfy in harmful, unintended ways. In banking this guards against an agent that, told to 'maximise collections' or 'reduce processing time', cuts corners on compliance or customer fairness. The spec captures the multi-dimensional objective: the goal, the constraints, the prohibited behaviours and the success criteria, so intent can be compared against what the agent actually does. To implement, require a signed intent spec to accompany each agent before it goes live, store it as version-controlled evidence, and run periodic intent-versus-behaviour audits that sample real agent traces and judge whether actions stayed aligned with the documented intent. There is no formula here; verification is audit-based. The threshold has two parts: 100% of metric-driven agents must carry a documented intent spec, and the intent-versus-behaviour audit pass rate must be at least 95%. Falling short - a missing spec or an audit pass rate under 95% - triggers remediation, blocking deployment or pausing the agent until alignment is restored.

Implementation notes

Use structured goal templates with human-readable intent declarations alongside measurable proxies. Run quarterly intent audits comparing observed behaviour to documented objectives. Employ multi-objective reward frameworks that resist proxy gaming. Test for Goodhart's Law vulnerabilities in all metric-driven agents.

Risks mitigated

4