> ## Documentation Index
> Fetch the complete documentation index at: https://agentconcepts.io/llms.txt
> Use this file to discover all available pages before exploring further.
# satisficing
> Strategic acceptance of a good-enough option rather than continued search for the optimum; the bounded-rationality response when full optimization is infeasible or its cost exceeds the marginal gain.
business computer-science economics psychology
# Satisficing
## Description
Strategic acceptance of a good-enough option rather than continued search for the optimum. Coined by Herbert Simon (portmanteau of "satisfy" + "suffice") as the bounded-rationality counterpart to optimization. The searcher sets an aspiration level — a threshold of acceptability — and accepts the first option that clears it, rather than enumerating and ranking the full option space.
The diagnostic move is the *aspiration level*: it's the structural element distinguishing satisficing from random first-pick (no bar) and from optimization (no stopping at first-over-bar). The aspiration level is itself a design choice — too high and search never terminates; too low and quality suffers; the calibration of the bar is where the concept's value lives.
Distinct from local-minimum: local-minimum is involuntary structural stuckness (every nearby move is worse, so the searcher can't escape); satisficing is voluntary search termination (the searcher chooses to stop because further search isn't worth the cost). Same observable outcome — no further improvement — from opposite mechanisms.
## Triggers
**User-initiated:** User describes a decision being made by "good enough" criteria, or asks about when to stop searching/iterating. Vocabulary cues: "satisfice," "good enough," "stopping rule," "aspiration level," "80/20," "diminishing returns," "ship it."
**Agent-initiated:** Agent notices that continuing search has marginal cost exceeding marginal value, and the question is what aspiration level justifies stopping. Candidate inference: "what's the bar; have we cleared it; is further search worth its cost?"
**Situation-shape signals:** Open-ended search problems with unbounded option spaces. Decisions under time pressure. Iterative improvement loops where each iteration's marginal value is decreasing. Any context where "the perfect is the enemy of the good" applies.
## Exclusions
* **Stakes are extreme** — when the cost of a suboptimal choice is catastrophic (medical diagnosis, safety-critical engineering), satisficing's "first acceptable" is too risky; full optimization or expert consultation is warranted.
* **Search space is small** — when you can enumerate all options cheaply, optimize rather than satisfice; satisficing is the bounded-rationality response, and the bound has to actually bite.
* **No meaningful aspiration level can be set** — for genuinely open-ended creative problems where "good enough" is the question being investigated, satisficing collapses into "pick anything."
* **Repeated/aggregate decisions** — a single satisficed decision is fine; a long series of satisficed decisions accumulates suboptimality. Repeated-game contexts may merit more careful per-decision optimization.
## Structure
## Relationships
* [local-minimum](/concepts/local-minimum) — contrast: local-minimum is involuntary-stuck; satisficing is voluntary-stop. Naming the difference matters because the responses differ (escape vs. confirm-bar-is-right).
* [asymmetric-gate](/concepts/asymmetric-gate) — the aspiration level is an asymmetric gate: options that meet the bar pass cheaply (accept-and-stop); options that don't get rejected and search continues. The asymmetry is the concept's productive feature.
* [doctrine](/concepts/doctrine) — a good satisficing decision is governed by a doctrine that encodes the aspiration level (with rationale and trigger) rather than ad-hoc "feels good enough."
* [cost-cascade](/concepts/cost-cascade) — satisficing is the early-stop in a cost-cascade: cheap first option succeeds, skip the expensive fallback search.
* [load-bearing](/concepts/load-bearing) — the aspiration level is load-bearing; remove it and you're either random-picking or optimizing, neither of which is the same primitive.
## Examples
Yelp 4+ stars and reasonably close beats hour-long search for the platonic best dinner.
accept the first implementation that passes the test bar; don't refactor toward the platonic optimum.
buying the first apartment that meets the must-haves rather than touring 30 more; the bar (price, size, location) is the aspiration level.
Gigerenzer and Selten's edited volume *Bounded Rationality: The Adaptive Toolbox* reframed Simon's bounded-rationality program away from "the way humans approximate optimal decisions under cognitive constraints" toward "the way humans use fast-and-frugal heuristics that exploit environmental structure to outperform optimization in real ecological contexts." The "adaptive toolbox" metaphor positions satisficing not as a fallback when optimization is infeasible but as a *first-class strategy* whose properties (low information cost, robustness to overfitting, ecological validity) often dominate the apparently-better optimizing alternatives.
The structural insight is that satisficing's apparent suboptimality compared to full optimization is measured against a counterfactual that often does not exist — full optimization requires information the agent does not have and cannot obtain, so the relevant comparison is not "satisfice vs optimize" but "satisfice well vs satisfice badly." The research program shifts attention to what makes a specific heuristic ecologically valid: the match between its computational shortcut and the structure of the environment it is deployed in.
**Inference**: When evaluating any decision-making system that uses simple rules ("take the first acceptable option," "default to last-known-good," "use the most-recent-N as the estimate"), the right comparison is not against an idealized full-optimization that nobody can actually run. It is against alternative simple rules calibrated to the same information cost. The "adaptive toolbox" framing makes the design question explicit: what is the right tool for *this* environment, given *these* costs of information?
accept the first offer above a salary/role bar rather than continuing to interview indefinitely; the bar trades off opportunity cost against acceptance speed.
training halts when validation loss is "good enough" or stops improving; the convergence threshold is the aspiration level. Distinct from convergence-as-optimum because the stop is by rule, not by exhaustion.
Simon's original case: corporate decisions aren't optimized but satisficed because the cost of full optimization exceeds the marginal value.
MVP / "good enough for users" instead of feature-complete; aspirations are explicit (what does v0 need to clear?).
Barry Schwartz's *The Paradox of Choice* popularized a personality-trait distinction supported by his and Sheena Iyengar's research: *maximizers* are agents who search for the best option in any decision and feel ongoing regret about unexplored alternatives; *satisficers* are agents who accept the first option above their aspiration level and disengage from further search. On standard well-being measures, satisficers score substantially better than maximizers, despite (in many domains) choosing objectively worse options. The disposition to maximize imposes search cost, attention cost, and regret cost that often exceeds the marginal value of a slightly-better outcome.
The structural insight is that the cost side of the optimization-vs-satisficing tradeoff includes psychological costs beyond information cost — the rumination overhead of unfinished search, the regret overhead of comparing against unchosen alternatives, and the attention overhead of staying open to revision after the decision. Satisficing's discipline includes the cognitive move of *closing* the decision once made.
**Inference**: For decisions where outcome differences are small relative to search and rumination costs (most consumer choices, most reversible commitments, most peer-level professional decisions), the satisficing posture has lower total cost than maximizing even when measured purely in the agent's own well-being. Designing decision processes (and product interfaces) that *enable* satisficing — clear good-enough thresholds, low-regret commitment mechanisms, friction against post-decision reopening — is the often-overlooked half of the bounded-rationality program.
Herbert Simon's 1956 *Psychological Review* paper "Rational choice and the structure of the environment" coined *satisficing* — a portmanteau of *satisfy* and *suffice* — and developed it as a normative model of decision-making for agents with bounded computational capacity in environments rich enough that full optimization is infeasible. The paper showed mathematically that an agent following a simple "search until the first option exceeds an aspiration level, then stop" rule could perform near-optimally in many naturalistic environments, *and* that the cognitive simplicity of the rule was structurally what made it possible to deploy at the agent's actual computational scale.
The contribution that earned Simon the 1978 Nobel Prize in Economics was the recognition that decision-quality is the product of decision-rule-quality *and* the agent's actual ability to execute the rule given finite resources. An expected-utility maximizer who cannot in practice compute the expected utility is a worse decision-maker than a satisficer who can actually execute their rule.
**Inference**: When designing decision procedures for agents (human or machine) operating under realistic resource constraints, the load-bearing question is not "what is the optimal decision rule?" but "what is the best decision rule the agent can actually execute end-to-end?" A simpler rule that runs to completion will routinely outperform a sophisticated rule that times out or quietly approximates. Simon's framework is the principled basis for accepting good-enough as the right answer, not as a compromise.
Simon's 1957 *Models of Man: Social and Rational* collected and extended his work on bounded rationality across organizational behaviour, administrative decision-making, and economic theory. The book's contribution was to take satisficing out of the laboratory and into descriptions of how real organizations make consequential decisions: hiring committees do not enumerate the global candidate pool, executive teams do not exhaustively compare strategies, regulators do not optimize across all rule variants. They search until they find an acceptable option, then stop — and the *aspiration level* (the threshold of acceptability) is itself adjusted based on the search experience.
The structural insight that recurs across the chapters is that organizations and individual agents share a common pattern: aspiration levels rise when search is easy and fall when search is hard, producing a self-calibrating mechanism that keeps the search cost roughly proportional to the available payoff. The aspiration-adjustment dynamic is what prevents satisficing from collapsing into "first thing I see" — the threshold tracks the structure of the environment.
**Inference**: When implementing any system that has a "good enough" threshold (acceptance criteria for code reviews, stopping conditions for search algorithms, quality bars for LLM generations), the threshold should not be a fixed constant. It should adjust based on observed difficulty: low-bar acceptance when the available options are mediocre, high-bar acceptance when the available options are excellent. The self-calibrating aspiration-level is the structural property that makes satisficing robust across environments.