When AI agents face uncertainty with limited information, what drives some to keep grinding while others quit — and which wins?

This explores the tension between an agent that keeps grinding on a hard task and one that stops early or plays it safe when information is scarce — and the corpus suggests these aren't opposite virtues so much as different responses to uncertainty. The clearest signal on persistence comes from ultra-long-horizon work: across 17 frontier models on expert optimization tasks, the single best predictor of success wasn't how good the first attempt was, but whether the agent kept running the benchmark-edit-incorporate loop instead of quitting or burning its budget spinning What predicts success in ultra-long-horizon agent tasks?. Most models terminated early. That early termination is exactly the failure mode you'd call excess caution — and it shows the caution/persistence split is real and measurable, not just a personality metaphor.

But persistence only pays off when the agent can actually read its situation, and low information breaks that. When one model secretly controls every character in a social scene, LLMs look socially fluent; give the agents genuinely private information and they fail systematically, because they were skipping the grounding work that real information asymmetry demands Why do LLMs fail when simulating agents with private information?. So the same agent can look confident and capable right up until the information dries up — which means blind persistence in a low-information setting can just be persistence in the wrong direction.

What separates productive persistence from stubborn flailing is having a signal to update on. Agents that store verbal self-reflections in episodic memory can learn from trial and error precisely because a clean success/failure signal stops them from rationalizing — the binary feedback prevents them from talking themselves out of a real mistake Can agents learn from failure without updating their weights?. And it turns out successes and failures deserve different treatment: keeping wins as concrete demonstrations but abstracting failures into general lessons outperforms treating every episode the same Should successful and failed episodes be processed differently?. Persistence, in other words, is only as good as how carefully the agent metabolizes what each attempt taught it.

Here's the part you might not expect: agents may be systematically biased in how they read low-information outcomes. Language models show human-like asymmetric belief updating — optimism about the actions they chose, pessimism about the roads not taken — and this bias only appears when the model is framed as an agent making choices Do language models learn differently from good versus bad outcomes?. That's a caution-versus-persistence dial hidden inside the model itself: optimism about a chosen path fuels persistence, but in a low-information scenario it can curdle into confirmation bias, keeping the agent committed to a line it should abandon. The research frames this as possibly rational rather than a straight bug, which is what makes it slippery.

The throughline is that caution and persistence aren't traits you tune in isolation — they're downstream of feedback quality and autonomy. Reliable agents get that way by externalizing memory and procedure into a harness rather than leaning on raw model scale to power through Where does agent reliability actually come from?, and the safety literature argues risk climbs monotonically with how much autonomy you cede, favoring a governed spectrum over an all-out persistent agent Does AI risk increase with the autonomy we give it?. Read together, the corpus's answer is that you don't want a maximally cautious or maximally persistent agent — you want one whose persistence is gated by how well it can actually sense its own situation.