[philosophy] [aquinas] Three-Phase Agent Loop: Counsel, Judgment, Command #271
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: Rockachopa/Timmy-time-dashboard#271
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Source: Thomas Aquinas, Summa Theologica II-II Q47 (Prudence, considered in itself), Q49 (The integral parts of prudence), and Q51 (The virtues connected with prudence). Dominican Province translation, 1920. Retrieved from https://www.newadvent.org/summa/
The Thomistic Action Architecture
Aquinas identifies three acts that constitute the full operation of practical reason in action:
He then distinguishes three subordinate virtues corresponding to these acts:
Key insight: "After judging aright we delay to execute or execute negligently or inordinately. Hence after the virtue which judges aright there is a further need of a final and principal virtue, which commands aright, and this is prudence." (Q51 A3, Reply 3)
The Principle
An agent's reasoning process is not a monolithic act but a three-phase deliberative cycle with categorically different failure modes at each stage:
These are separate virtues because they can fail independently. You can be good at research but terrible at judgment. You can judge well but fail to execute. And critically, command is the chief act — the one closest to the end — meaning that an agent optimized for counsel or judgment but weak in command is fundamentally defective.
Application to Agent Architecture
Modern AI agent loops conflate these three acts into a single "reasoning" step. The agent generates tool calls (counsel + judgment + command simultaneously), executes them, observes results, and repeats. This architecture makes three distinct failure modes invisible:
Counsel failure (bad research):
Judgment failure (bad evaluation):
Command failure (bad execution):
Current agent architectures treat all three as "reasoning errors" and respond the same way: generate more tokens, try again. Aquinas shows this is wrong. The virtues are separate because the corrections are separate.
Proposed Action
Implement a three-phase agent loop with explicit separation and phase-specific failure detection:
Phase 1: Counsel (euboulia)
Phase 2: Judgment (synesis/gnome)
Phase 3: Command (prudence proper)
Implementation
Add a
deliberation_modeparameter toAIAgent.run_conversation():"monolithic"(default, current behavior) — single-phase reasoning"thomistic"— three-phase loop with explicit counsel/judgment/command separationIn Thomistic mode:
Phase-specific failure detection:
This creates an observable separation between "I didn't think of the right approach" (counsel), "I chose the wrong approach" (judgment), and "I failed to execute the right approach" (command) — three categorically different agent pathologies that currently all look like "the agent is bad at reasoning."
Concrete next step: Prototype
deliberation_mode="thomistic"inrun_agent.py, test on complex multi-step tasks where current single-phase loops thrash, and measure whether explicit phase separation reduces iteration count and improves success rate.Consolidated into #300 (The Few Seeds). Philosophy proposals dissolved into 3 seed principles. Closing as part of deep triage.