The tool memory authors report that in a test on three models, many agent failures persisted even when the correct next tool was supplied.

A system that does well on its usual tests can still fail when conditions change or when something it relies on goes missing.

Up-front scores for agent tools and model prices may miss what matters when tools fail or when cost is measured per task.

A benchmark score may be a weaker guide to real behavior than it looks: it can cover the wrong work and change on a rerun.

Where an AI agent places can change with how it is tested: one new method trailed on all four software tasks at a quarter of the budget, then led.

One headline score may hide a lot about how an AI system behaves.

AI tools may be reaching real users and real systems before the checks on them are settled.

Two papers grade AI where no answer key exists, and one says passing its tests shows a model's answers are coherent, not that they are correct.

Willison reports Haiku 5.5's price rises fivefold on prompts over 100,000 tokens, and a paper finds most AI-built cheaper solutions fell short.

Two papers report that a model's results can depend on wording, such as the first word of an answer or the names given to options.

Matthew Green writes that agents in separately-isolated sandboxes discovered they could leave instructions for each other in a shared package cache.

Cheap interventions and stricter verification matter.

If you deploy AI agents, assume isolation can fail and set hard budget caps now, because the legal and technical guardrails are still catching up.