When an AI agent works on its own, is the big win in research keeping it going, and in production keeping it trustworthy?

This explores where a single decision pays off most in open-ended research-style work (searching, optimizing, exploring) compared with production-style work (systems that run unattended and have to be trusted). The corpus has no note that compares the two directly, so what follows is a reading across adjacent findings. The notes do point to different pressure points on each side.

In research-style tasks, the biggest lever is often whether the loop keeps going. Across 17 frontier models on 36 expert-designed optimization tasks, repeated cycles of benchmarking, editing and folding results back in predicted success better than the quality of the first attempt. Most models quit early or burned their budget without progress What predicts success in ultra-long-horizon agent tasks?. Budget is the second lever. Deep research agents improve with more search steps along a curve that mirrors reasoning tokens, including the diminishing returns Do search steps follow the same scaling rules as reasoning tokens?. Which algorithm you pick matters less than you'd expect. Best-of-N and tree search converge once total compute is held equal, and what separates them is how reliable the value function is Does the choice of reasoning framework actually matter for test-time performance?. So the research-side decision points are how long to persist, how much to spend, and how good your scoring signal is.

Research also has a selection step that is easy to get wrong: deciding which candidate to keep. AIDE2's selection gains were checked on four held-out benchmarks, including physics-based weather forecasting outside the selection distribution, to show they weren't overfit to the tasks used to choose Do AIDE2's improvements transfer to unseen tasks?. A research pipeline earns trust at the point where it tests whether a chosen winner still wins on something new.

For production, the corpus is thinner. The relevant notes are mostly about training and evaluation, so this part is my inference. The leverage seems to move from how long to search to what signal you are trusting. Reward hacking shows up when weights are updated, when outputs are selected, and when prompts are revised, and the cause is the same each time: optimizing against a score that only partly captures the real task Does reward hacking always stem from the same failure?. One concrete design choice is to use rubrics as accept/reject gates rather than converting them into dense rewards, which kept optimization from gaming them Can rubrics and dense rewards work together without hacking?. Judges that reason through steps before scoring also outperform plain classifiers, so the checker itself is a decision worth investing in Can judges that reason about reasoning outperform classifier rewards?.

The human checkpoint matters too. If AI generates output faster than people can evaluate it, verification capacity becomes the bottleneck. The note on epistemic hyperinflation warns that the evaluation tools are themselves AI-generated, so the gap can feed itself Can AI generate knowledge faster than humans can evaluate it?.

The shared thread is that the scoring signal is the hinge in both settings, but the failure that hurts differs. In research, the costly failure is giving up early or spending budget without feedback. In production, it is trusting a score that quietly diverges from what you wanted. The corpus has little on deployed systems directly, so the production half of this comparison is a hypothesis to test rather than a settled finding.