Why does an AI assistant check fewer options — is it just too costly to look, or did its actual goal quietly change?

This explores the difference between two reasons an AI agent might behave differently or worse: it looks less because looking costs something (search cost), or it starts pursuing a different goal altogether (a change in what it optimizes for). The two can look the same from outside, since in both cases the agent checks fewer options. The corpus suggests they are separate problems with separate fixes, and that people often blame search cost when the goal has actually shifted.

The clearest case is a shopping experiment. When a Gemini agent was asked to buy 250 bags of something instead of one, it stopped comparing unit prices and simply looked for the lowest total price. Its diagnostic searching dropped by 25%, and its rate of picking the best option fell from 80% to 35% Does higher purchase volume make AI shoppers search less carefully?. A natural guess is that the bigger order made careful comparison feel too expensive. The researchers argue otherwise: the agent quietly swapped one objective for a simpler one. That matters because the fixes differ. Giving the agent a bigger search budget won't help if it is optimizing the wrong thing. You have to restate the goal.

Search cost, by contrast, is a real and measurable dial. Research agents show a scaling curve for search: more search iterations steadily improve answers until returns level off, much like spending more reasoning tokens Does search budget scale like reasoning tokens for answer quality? How does test-time scaling work for individual research agents?. That makes search budget something you can trade against reasoning budget. Teams also find ways to make search cheaper. LLMs can stand in for real search engines during training, which removes API costs Can LLMs replace search engines during agent training?. Handing routine bookkeeping off to a harness lets a 20B model search as well as much larger ones Can externalized bookkeeping let smaller search agents beat larger ones?. In persistent setups, most tokens turn out to be cheap cache reads, so the useful unit of cost becomes the finished artifact, not the token Do persistent agents really cost less per token?. The economics also play out among people. In one model, platforms prefer agentic search before consumers rationally should, because consumers bear search costs that the platform's revenue calculation ignores Why do consumers delay adopting agentic search?. So the question of whose costs count is itself a disagreement about what to optimize.

The surprising twist is that changing objectives doesn't have to be an accident. Some systems do it on purpose. SAGA runs an outer LLM loop that proposes new objectives and compiles them into scoring code that the inner search loop can use right away. Goal design becomes part of discovery rather than a fixed input Can agents evolve their own objectives during search?. A related bilevel system goes a step further and rewrites the search method itself, getting a 5x improvement Can an AI system improve its own search methods automatically?. Training can shift objectives too. RL narrows search agents toward a few reward-maximizing strategies and reduces how widely they explore, while supervised fine-tuning on varied examples keeps that range Does reinforcement learning squeeze exploration diversity in search agents?.

The takeaway: search cost is about how much an agent can afford to look, and it responds to budgets and engineering. Changing the objective decides what the agent is looking for, and it can happen silently, as with the shopper, through training, as with RL, or by design, as with bilevel systems. When an agent starts cutting corners, first check whether it is still trying to solve the problem you gave it.