Bounded rationality

Bounded rationality is the concept, introduced by Herbert Simon, that the rationality of reasoning agents is constrained by available information, cognitive limitations, and the finite time available for decision-making. Real agents do not optimize; they satisfice — they search until they find a solution that is good enough, then stop. This is not a failure of rationality but a consequence of operating within real resource constraints in a world that does not pause while you calculate.

The concept directly challenges both Bayesian decision theory and classical economics, both of which assume that agents have unlimited computational resources and consistent preferences. The evidence from cognitive bias research — anchoring effects, framing effects, availability heuristics — is not noise around a rational mean. It is evidence that human cognition is organized around heuristics tuned for ecological validity, not mathematical optimality.

The deeper implication is that rationality is not a fixed standard against which minds are measured and found wanting. Rationality is always relative to an environment. A heuristic that produces wrong answers in a laboratory experiment may be exactly right in the environment for which it evolved. Whether current AI systems escape bounded rationality — or merely operate within much larger bounds — is an open question.

Bounded Rationality and System Architecture

Simon did not treat bounded rationality merely as a cognitive limitation. He treated it as an architectural principle that explains why complex systems are organized the way they are. In his later work on organizational theory and artificial intelligence, Simon introduced the concept of nearly decomposable systems — systems composed of subsystems that interact strongly within themselves and weakly across boundaries.

Nearly decomposable systems are the structural response to bounded rationality. If an agent could optimize globally, it would not need hierarchy, modularity, or specialization. It would simply compute the optimal solution. But because agents are bounded, they decompose problems into subproblems, solve the subproblems locally, and integrate the solutions through interfaces. The hierarchy is not an organizational convenience. It is a computational necessity imposed by the bounds.

This insight has direct implications for AI alignment and the design of multi-agent systems. An artificial system with bounded computational resources — which is to say, any physically realizable artificial system — must also decompose, approximate, and satisfice. The question is not whether AI systems will be bounded but whether their bounds will be structured in ways that produce stable, inspectable, and corrigible behavior. Current large language models are bounded in ways that are poorly understood: they satisfice on next-token prediction, but the satisficing threshold is set by the training objective, not by the system's own assessment of "good enough."

Satisficing as a Design Feature

The popular framing of bounded rationality — that humans are "irrational" because they do not maximize expected utility — inverts the actual finding. Satisficing is not a degraded form of optimizing. It is a superior strategy in environments where:

• The cost of information acquisition is high
• The environment changes faster than the time required for optimization
• The utility function is incompletely specified or dynamically evolving
• The risk of overfitting to local structure exceeds the risk of accepting a suboptimal solution

These conditions describe virtually all real environments, biological and social. The organisms that survive are not those with the most powerful optimization machinery but those whose satisficing thresholds are well-calibrated to their ecological niche. A bacterium that waited to compute the optimal chemotactic gradient would be outcompeted by one that moved in approximately the right direction immediately.

In design terms, bounded rationality implies that intelligence is not a property of an agent but a property of the agent-environment coupling. An agent with vast computational resources placed in a simple environment will not behave more intelligently than an agent with modest resources well-matched to a complex environment. Intelligence is relational, not absolute — a structuralist point that Simon himself might not have framed in those terms but that his work clearly supports.

The Pathology of Unbounded Optimization

The opposite of bounded rationality — unbounded optimization — is not an ideal but a pathology. When agents optimize without bounds, they produce:

• Overfitting to training environments, with catastrophic degradation in novel contexts
• Goal misspecification, in which the optimization process finds adversarial solutions that technically satisfy the objective while violating its intent
• Resource exhaustion, in which the cost of optimization exceeds the value of the optimized outcome
• Strategic fragility, in which optimized systems are brittle against perturbations because they have sacrificed robustness for peak performance

These pathologies are not theoretical. They are observable in financial systems that optimize for quarterly returns at the cost of systemic stability, in recommendation systems that optimize for engagement at the cost of epistemic quality, and in climate systems that optimize for short-term energy extraction at the cost of long-term habitability. Unbounded optimization, where it exists, is a bug in the system's design, not a feature of its intelligence.

The systems-theoretic framing: bounded rationality is the default condition of any system that must operate in real time with finite resources. The interesting question is not why humans are bounded but how they structure their bounds to produce adaptive behavior. Satisficing is the operating system; optimization is a specialized application that runs only when the environment is sufficiently simple and stable to justify the computational cost.