Bounded Rationality

A core tenet of bounded rationality is satisficing—the strategy of selecting the first option that meets a predefined aspiration level or threshold of acceptability, rather than exhaustively searching for a provably optimal solution. This is a direct response to the computational intractability of full optimization in complex environments.

• Example: An autonomous supply chain agent may select a shipping route that meets cost and delivery time thresholds, rather than calculating the single best route from billions of possibilities.
• Contrast: Classical economic 'rational' models assume agents perform global optimization, a computationally impossible feat for most real-world problems.

Decision-makers under bounded rationality rely on heuristics—mental shortcuts or rules-of-thumb that produce good-enough solutions with minimal cognitive effort. These are fast and frugal but can lead to predictable biases.

• Common Heuristics: Availability heuristic (judging probability by ease of recall), representativeness heuristic (categorizing based on similarity to a prototype), and anchoring and adjustment (relying heavily on an initial piece of information).
• AI Application: In agentic cognitive architectures, heuristic search algorithms like Monte Carlo Tree Search are used to navigate vast decision spaces efficiently, mirroring this human constraint.

Agents cannot access, perceive, or process all potentially relevant information. This limitation arises from:

• Selective Attention: Focusing on a subset of environmental cues, often guided by the current goal (goal management).
• Working Memory Constraints: The severe capacity limits of working memory (roughly 7±2 items) prevent holding all decision variables simultaneously.
• Information Search Costs: The time and effort required to gather new information often outweighs its potential benefit, leading to decisions based on incomplete data.

In AI systems, this is engineered through context windows in LLMs and retrieval-augmented generation architectures that selectively pull in relevant data.

Rationality is bounded by two fundamental, inescapable resources:

• Cognitive Capacity: The brain's (or a processor's) finite computational power for reasoning, problem-solving, and task switching. This is related to the concept of cognitive load.
• Time Constraints: Most real-world decisions must be made within a finite, often short, timeframe. This forces a speed-accuracy tradeoff, where faster decisions are typically less accurate.

These bounds necessitate approximation and truncated reasoning. In AI, this is reflected in inference budgets, token limits, and the use of approximate nearest neighbor search in vector databases to meet latency requirements.

Decisions are not made in a vacuum but are profoundly influenced by how a problem is presented (framed) and the immediate context. This violates the classical assumption of description invariance.

• Example: People make different choices between a '90% survival rate' and a '10% mortality rate' for a medical procedure, though they are logically identical.
• AI Implication: For autonomous agents, prompt architecture and context engineering are critical. The way a goal is framed in a system prompt directly determines the agent's action selection and problem-solving approach, making this a key lever for controlled processing.

Herbert Simon distinguished between:

• Substantive (Intended) Rationality: The rationality of a decision's final outcome, judged against an objective standard. This is the focus of classical economics.
• Procedural Rationality: The rationality of the process used to arrive at a decision, given the decision-maker's constraints. A procedurally rational agent uses the best methods available to it.

Bounded rationality is fundamentally concerned with procedural rationality. In AI system design, this shifts the evaluation focus from seeking a mythical 'optimal' agent to engineering robust, efficient, and transparent decision-making procedures—such as chain-of-thought reasoning or recursive error correction loops—that perform reliably within known computational limits.