Four studies today put distance between how agents score and how they behave

A state-of-the-art review of AI-assisted coding collected field experiments reporting 26% more tasks per week, randomized trials reporting a 19% slowdown, and team telemetry showing code review time up 441%. That spread is not noise about how good the tools are, it is evidence that the three are not measuring the same stage of the work. Two smaller results land in the same place from different directions: rewritten error messages that feel clearer without moving fix rates, and occupational bias that reads clean at the output while the model's internal estimate of user expertise still changes the answer. The consequence is practical, a passing evaluation licenses less than it used to, and the burden moves toward measurements taken under deployment conditions. Read: The review puts field experiments at 26% more tasks per week, randomized trials at a 19% slowdown, and team telemetry at 441% more code review time. These disagree because they stop counting at different points: work accepted, work completed, and the downstream cost of integrating what was accepted. Read: LLM-rewritten interpreter errors scored better on perceived clarity and cognitive load while fix rate, attempt count and time-to-fix did not move. Perceived improvement is cheap to measure and easy to ship, which is the reason it should not stand in for a performance claim. Read: Established test and evaluation practice assumes a system is specifiable, stable, composable and supervisable. Agentic behavior weakens all four at once, so a result describes the conditions of the test rather than the system that ships. Read: One result shows the block an agent needs evicted under a token budget before retrieval runs at all. Another shows retrieval serving stale facts about a third of the time when a value changes mid-session, with an LLM reranker failing to correct it. Both point at tracking supersession and dependencies rather than better ranking. Read: OpenAI and AWS report Terra completing Terminal-Bench 2.1 tasks at roughly 82% lower cost, with Sol and Luna also selectable. The claim is cost per solved task rather than cost per token, which is the right denominator and the harder one to reproduce on your own workload. Read: An attacker-controlled bug report can steer both the patch an agent writes and the description attached to it, so the reviewer reads text the attacker influenced. Human review stops being a control once the input to review is part of the attack surface. Read: How a skill is described in the system prompt, not just which skills are present, decides what the planner reaches for. An 8,135-trial study points the same way, finding skills work by anchoring procedure rather than supplying missing knowledge, so runbooks and domain routing come before catalog size.