We’re trying to make AI agents safe by giving them rules: guardrails, policies, allow-lists.
Aristotle saw the limit of this 2,300 years ago. Rules cover the cases you anticipated. Practical wisdom (phronesis, φρόνησις) is the ability to act well in the cases you didn’t.
An agent that follows every rule can still do the wrong thing: technically compliant, contextually absurd. A human expert doesn’t avoid that with more rules. They avoid it by understanding what the rules are for.
We can’t give agents phronesis yet. We can stop pretending that rule coverage is the same as judgement, and design accordingly:
- narrow autonomy where context matters,
- escalation instead of improvisation,
- humans placed where judgement is the actual work.
More rules make a more obedient agent. They don’t make a wiser one.