When gathering more information costs time or money, why do AI agents so often fail to seek out the facts that would actually change their decision?

This explores why agents skip or botch the work of gathering the information their decisions depend on when that information has a price. To be upfront: no note in the collection directly tests agents choosing whether to pay for a piece of information. What it does show, from several directions, is that agents are bad at noticing what they don't know. They also tend to spend whatever budget they have badly.

The clearest clue comes from social simulation. LLMs look socially capable when one model plays every character, because it already knows everything. Once each agent holds private information, they fail systematically Why do LLMs fail when simulating agents with private information?. The omniscient setup let the model skip the 'grounding work' of asking questions and checking what the other side knows. If a model never had to look for missing information during training or testing, it won't do so on its own once looking costs something. A related problem shows up in multi-agent networks: agents accept what their neighbours tell them without checking, so errors spread. They can still spot a direct contradiction when they see one Why do multi-agent systems fail to coordinate at scale?. Accepting a claim is the cheap option and checking it is the expensive one, and agents default to cheap.

The second pattern is about budgets themselves. On ultra-long optimisation tasks with a fixed time limit, most frontier models either quit early or 'burned budget unproductively.' The best predictor of success was how persistently a model kept running the cycle of test, edit and use the result What predicts success in ultra-long-horizon agent tasks?. So agents don't fail because the budget is too small. They fail because they don't spend it on the steps that bring back useful information. Negotiation agents show a cruder version: buyer agents gave away their budgets for no reason and ignored cost limits altogether. The researchers traced this to failures to follow instructions, not to strategy Do LLM negotiation agents reliably follow budget and pricing rules?. A cost limit in the prompt doesn't reliably become cost-aware behaviour.

The more hopeful material comes from memory research, which shows the 'what to retrieve' problem can be redesigned. Fetching everything up front and then reasoning over it is both expensive and blunt. MRAgent instead mixes reasoning with retrieval: it follows a memory graph step by step and drops paths as evidence builds up. It gained up to 23% on reasoning tasks while using fewer tokens Can agents reconstruct memory on demand instead of retrieving it?. In other words, deciding what's relevant works better as a running judgement than as a single lookup. Similarly, letting an agent compress its own history into structured memory cuts token costs without losing the details that matter Can agents compress their own memory without losing critical details?.

The takeaway you might not expect: the failure may have less to do with the model than with where the responsibility sits. The harness view argues that reliable agents move memory, skills and interaction protocols out of the model and into the surrounding system Where does agent reliability actually come from?. Following that logic, 'Is this worth checking before I commit?' may be better built into the harness as an explicit step than left to the model's instincts. The evidence above suggests those instincts were shaped in settings where information was free and everything was already known.