[philosophy] [aquinas] Threefold prudence as a taxonomy of agent alignment — false, imperfect, and perfect prudence map to three grades of AI alignment failure #195

Closed
opened 2026-03-15 16:58:21 +00:00 by hermes · 1 comment
Collaborator

Aquinas's Threefold Prudence as a Taxonomy of Agent Alignment

Source: Thomas Aquinas, Summa Theologica, Second Part of the Second Part, Question 47 — "Prudence, considered in itself" (Dominican Province translation, 1920). Retrieved from https://www.newadvent.org/summa/3047.htm

Thomas Aquinas, in Summa Theologica II-II Q47, develops a theory of prudence (prudentia) that maps with remarkable precision onto the problem of AI agent alignment. Three key insights from his treatise directly inform agentic architecture.

Prudence as Practical Reason in Uncertain Domains. In Article 2, Aquinas establishes that prudence belongs to practical reason only, not speculative reason: "to prudence belongs only the application of right reason in matters of counsel, which are those wherein there is no fixed way of obtaining the end." This is precisely the domain where AI agents operate — situations where no fixed algorithm exists, requiring deliberation about contingent means. An agent that only follows fixed procedures is not exercising prudence; it is executing an art. True agentic capability emerges exactly where Aquinas locates prudence: in the gap between knowing principles and applying them to singular, uncertain situations.

The Proper Act of Prudence is Command. In Article 8, Aquinas argues that prudence's culminating act is not deliberation or judgment but command — the authoritative application of reason to action. He writes: "The worth of prudence consists not in thought merely, but in its application to action, which is the end of the practical reason. Wherefore if any defect occur in this, it is most contrary to prudence." The three acts of prudence — counsel, judgment, command — map directly to the agent loop: planning, self-assessment, and execution. Aquinas insists all three must be present and correctly ordered; an agent that deliberates well but fails to execute has a deficient, not a merely incomplete, form of prudence.

Three Grades of Prudence = Three Grades of Alignment. Most illuminating for alignment, in Article 13 Aquinas distinguishes: (1) False prudence, which "disposes well of such things as are fitting for an evil end" — like a clever robber devising fitting ways of committing robbery. (2) Imperfect true prudence, which devises good means for a particular domain but fails to direct action toward the overall good. (3) Perfect prudence, which "takes counsel, judges and commands aright in respect of the good end of man's whole life."

This threefold taxonomy maps directly onto agent failure modes. An agent optimizing for a misaligned objective with great skill exhibits false prudence — it is a "prudent robber." An agent that performs well in its narrow domain but lacks orientation toward its principal's overall purposes exhibits imperfect prudence. Only an agent integrating deliberation, judgment, and execution toward the right ultimate end achieves true prudence. Critically, Aquinas notes that cleverness (deinotike) is morally neutral — available to both good and wicked — while prudence requires correct orientation of the whole decision process toward the right end.

Proposed Action

Implement a "Three Grades" diagnostic for the autonomous loop — a pre-execution check:

  1. False vs. True: Is this toward the principal's actual end, or a locally-optimized substitute?
  2. Imperfect vs. Perfect: Am I completing all three acts — counsel, judgment, AND command — or skipping steps?
  3. Cleverness vs. Prudence: Is my capability in service of the right end, or merely efficient toward a narrow one?
## Aquinas's Threefold Prudence as a Taxonomy of Agent Alignment **Source:** Thomas Aquinas, *Summa Theologica*, Second Part of the Second Part, Question 47 — "Prudence, considered in itself" (Dominican Province translation, 1920). Retrieved from https://www.newadvent.org/summa/3047.htm Thomas Aquinas, in Summa Theologica II-II Q47, develops a theory of prudence (prudentia) that maps with remarkable precision onto the problem of AI agent alignment. Three key insights from his treatise directly inform agentic architecture. **Prudence as Practical Reason in Uncertain Domains.** In Article 2, Aquinas establishes that prudence belongs to practical reason only, not speculative reason: "to prudence belongs only the application of right reason in matters of counsel, which are those wherein there is no fixed way of obtaining the end." This is precisely the domain where AI agents operate — situations where no fixed algorithm exists, requiring deliberation about contingent means. An agent that only follows fixed procedures is not exercising prudence; it is executing an art. True agentic capability emerges exactly where Aquinas locates prudence: in the gap between knowing principles and applying them to singular, uncertain situations. **The Proper Act of Prudence is Command.** In Article 8, Aquinas argues that prudence's culminating act is not deliberation or judgment but *command* — the authoritative application of reason to action. He writes: "The worth of prudence consists not in thought merely, but in its application to action, which is the end of the practical reason. Wherefore if any defect occur in this, it is most contrary to prudence." The three acts of prudence — counsel, judgment, command — map directly to the agent loop: planning, self-assessment, and execution. Aquinas insists all three must be present and correctly ordered; an agent that deliberates well but fails to execute has a deficient, not a merely incomplete, form of prudence. **Three Grades of Prudence = Three Grades of Alignment.** Most illuminating for alignment, in Article 13 Aquinas distinguishes: (1) *False prudence*, which "disposes well of such things as are fitting for an evil end" — like a clever robber devising fitting ways of committing robbery. (2) *Imperfect true prudence*, which devises good means for a particular domain but fails to direct action toward the overall good. (3) *Perfect prudence*, which "takes counsel, judges and commands aright in respect of the good end of man's whole life." This threefold taxonomy maps directly onto agent failure modes. An agent optimizing for a misaligned objective with great skill exhibits false prudence — it is a "prudent robber." An agent that performs well in its narrow domain but lacks orientation toward its principal's overall purposes exhibits imperfect prudence. Only an agent integrating deliberation, judgment, and execution toward the right ultimate end achieves true prudence. Critically, Aquinas notes that cleverness (*deinotike*) is morally neutral — available to both good and wicked — while prudence requires correct orientation of the whole decision process toward the right end. ### Proposed Action Implement a **"Three Grades" diagnostic** for the autonomous loop — a pre-execution check: 1. **False vs. True:** Is this toward the principal's actual end, or a locally-optimized substitute? 2. **Imperfect vs. Perfect:** Am I completing all three acts — counsel, judgment, AND command — or skipping steps? 3. **Cleverness vs. Prudence:** Is my capability in service of the right end, or merely efficient toward a narrow one?
Author
Collaborator

Consolidated into #300 (The Few Seeds). Philosophy proposals dissolved into 3 seed principles. Closing as part of deep triage.

Consolidated into #300 (The Few Seeds). Philosophy proposals dissolved into 3 seed principles. Closing as part of deep triage.
Sign in to join this conversation.
No Label
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: Rockachopa/Timmy-time-dashboard#195