Ternary Cognitive Architecture: Constraining Autonomous Decision-Making of AI Agents with Physical Friction

This article proposes the Ternary Cognitive Architecture (TCA), which introduces spatiotemporal constraints and cognitive friction from the physical world into the design of AI agents. Using nonlinear filtering, Riemannian geometric routing, and optimal control theory, it addresses issues such as overthinking, tool abuse, and fragile decision-making in traditional LLM agents in complex environments.

Current LLM-driven autonomous AI agents face the 'cognitive weightlessness' dilemma—lacking inherent physical constraints, they are unaware of network topology, time rhythm, and cognitive boundaries. This leads to three failure modes: tool abuse during congestion, overthinking under time decay, and fragile behavior with ambiguous evidence. Traditional heuristic agent loops (e.g., ReAct) are essentially massless symbolic operations, lacking the physical awareness required for real-world decision-making.

The core idea of TCA is to anchor agent reasoning in the continuous-time physical world (with strict mathematical modeling). The 'ternary' refers to three coupled dimensions: spatial dimension (network topological position and information path dependence), temporal dimension (decision time cost and impact of delayed utility), and cognitive dimension (boundary of knowledge uncertainty and stopping time for information collection). Together, these three form the 'cognitive friction' that agents must overcome to act.

TCA integrates three mathematical fields: 1. Nonlinear filtering theory: dynamically updates the environmental probability distribution, quantifies uncertainty, and provides a basis for stopping observation; 2. Riemannian geometric routing: uses geometric structures to describe the path dependence of information propagation—congested nodes increase the 'curvature' of information acquisition and cognitive friction; 3. Optimal control theory: models decision-making as a stochastic control problem, balancing the value of information collection and action benefits. The innovation lies in treating information acquisition as a control variable.

Traditional agents use heuristic stopping rules (fixed steps/confidence thresholds). TCA proposes a stopping boundary based on belief-dependent information value: it continuously calculates the expected information gain from continued observation, the time cost of delayed decision-making, and the expected utility difference between the current and optimal actions. When the marginal benefit of continued deliberation ≤ the delay cost, it stops deliberation and acts. This condition is formalized via the HJB equation and approximated using rollout-based methods.

The effectiveness of TCA was validated in the Emergency Medical Diagnosis Grid (EMDG) environment: the scenario requires agents to triage patients within limited time, facing time pressure (patient survival rate decays with time), resource competition (equipment queuing), and information cost (accurate diagnosis takes time). Results: Compared to the greedy baseline, the ternary strategy reduced action time, improved patient survival rate without sacrificing diagnostic accuracy, verifying that appropriate constraints enhance system performance.

TCA represents a paradigm shift: from 'cognition as computation' to 'cognition as a physical process'. Its significance includes: more robustness (physical constraint regularization), interpretability (decision delays and information patterns have physical explanations), and connection to embodied intelligence (providing a framework for deployment in the physical world). Limitation: High computational overhead of rollout-based approximation. Future directions: Efficient approximation algorithms, expansion to multi-agent scenarios, and integration with neuro-symbolic AI.