Cost-theoretic vs. panoramic approaches to resource rationality, after Icard (2025)

Given a class $\Pi$ of possible “agent programs” that implement a particular strategy, @2025icardResource defines two broad approaches to resource-rational analysis—that is, assessing an agent with limited resources—that trade off conceptual clarity and ease of application in opposite ways.

Name of approach	Definition of cost	Strengths	Weaknesses
Cost-theoretic	For each program $\pi \in \Pi$, estimate a cost function $C(\pi )$ (often the expectation of a random variable) for performing operations and updating internal states as stipulated by $\pi$. Different programs are compared by both how well they perform the task (i.e., take the sequence of actions that most rationally responds to information from the environment) and how efficiently they mitigate resource costs.	Easy to apply: need only specify a task (e.g., a specific decision problem) and plausible costs associated with different strategies for solving that task. Programs are assessed with a straightforward linear combination: $V(\pi )=U(\pi )-C(\pi )$.	Unclear what costs are, how they should be measured, and why they are commensurable with other measures of performance, e.g., added and subgracted with utility.
Panoramic	Instead of analyzing costs of specific operations, restrict $\Pi$ to programs satisfying a fixed “resource budget” threshold, and ignore differences in cost for programs in $\Pi$ (i.e., assume $v(\pi )=u(\pi )$ for all $\pi \in \Pi$).	Clear conceptual interpretation: which way of using a fixed set of resources is most conducive to a particular well-defined goal?	Need to comprehensively state the tasks faced by the agent and the space of all strategies the agent could be using, given available resources. Requires functional understanding of the agent, e.g., successfully applied to low-level perception.