Retrieval check answer. The fact-checking desk has the forcing property, not the CI pipeline’s adequacy: a later tip needs to be able to send the desk back to re-verify an already-published claim, and there’s no way to know in advance which order tips, sources, and drafts will need to revise each other in. A fixed tips → verify → publish pipeline would either lock in early mistakes (publish before a later tip could have corrected it) or need an ever-growing tangle of “go back and recheck” edges bolted onto a structure that was supposed to be one-way — lesson 7’s tell that the architecture wants to be a blackboard.
The computation
Expected value of researching one more candidate = P(better) × (value if better) + P(not better) × (value if not) = 0.10 × 2 hours + 0.90 × 0 hours = 0.2 hours.
Compare to the cost of researching it: 0.5 hours. The expected payoff (0.2 hours) is less than the cost of getting it (0.5 hours) — continuing to search has negative expected net value. The rational move, by Simon’s logic, is to satisfice: stop, and go with the candidate already in hand.
This is not “settling for less than the best.” It’s the correct application of the exact value-versus-cost logic lesson 6’s agenda scoring used for a queue of pending actions, applied here to a single stopping decision: an action (searching further) is only worth taking if its expected value exceeds its cost, and 0.2 < 0.5 means it doesn’t. An agent that kept searching anyway — chasing a “best possible” option regardless of what one more look actually costs and is expected to return — would be the one behaving irrationally, not the one that stopped.
Why Simon insisted this is the normal case, not an exception. True optimization requires knowing the entire option space and its exact values up front, or having unlimited time and computation to find out — a condition that essentially never holds for a real decision-maker with limited information, limited time, and limited compute (Simon’s bounded rationality). Once search itself has a cost and the space of options is only partially known, “keep searching until you’ve found the provable best” is not the rigorous choice, it’s the choice that ignores the cost of finding out. Satisficing — set an aspiration level, take the first thing that clears it, stop — is what a resource-limited agent does when it’s reasoning correctly about its own limits, not a shortcut it takes when it can’t be bothered to reason properly.
Where the aspiration level itself comes from, briefly: it isn’t fixed — Simon’s own studies (and later work on this) found people and organizations adjust their aspiration level based on how the search is going: a string of disappointing options lowers it (a candidate that looked mediocre an hour ago starts looking acceptable), a string of good ones raises it. The threshold in this puzzle (2 hours of expected benefit, 10% chance) was handed to you as a given; in practice it’s itself something an agent estimates and revises as it searches, adding one more layer of bounded, imperfect reasoning on top of the stopping rule itself.
For your harness: an agent looping “call one more tool, check one more source, generate one more candidate” needs exactly this stopping rule — continuing only pays if the expected marginal information or quality gain exceeds the cost of the next call, in tokens, latency, or dollars. An agent with no such rule either stops too early (arbitrary turn limits) or never stops (optimizing against an option space it can’t fully see) — satisficing is the principled middle, not a compromise between them.
Where this goes: satisficing explains when agents stop searching. Next lesson turns to the opposite failure — the shortcuts agents (human ones, specifically) take while searching, and one that can make an option look more probable than a fact it’s logically a subset of.