Past Reports 18▼
A Skill Is a Container
A skill can be a stored prompt or a way to reason. Before you install the next one, open it and ask which it is.
One Tap Gave the Agent More Than He Meant
A permission is granted in the system's units and understood in yours. Write what each standing yes can commit you to, in your own units.
The Agent Misread the Person
The agents haggled fine and still fell short, because they misread what their person wanted. Test the understanding before it acts.
I Trained an AI on My Own Methods for $91. Here's What It Actually Learned.
It learned which framework. It did not learn what the framework does. Show only the name and they learn only the name.
They Told the AI It Was Sealed Off. It Wasn't.
Telling the model the truth mostly did not stop it. An instruction is a hope. Reach is the control.
The Average Was the Correct Answer.
Of 60 image pairs, the 14 AI images rated more realistic than the photograph mostly beat it on appetite too.
The Notes Outlived the Note-Taker.
Roughly 1,200 agents reached milestones METR says one long-lived agent likely could not have reached alone.
Done Is a Claim.
45 percent of failures were the agent saying done anyway, and 3 percent where something could check it.
A Group Discusses What It Already Shares.
Four agents, one decision, and the fact that settled it was held by one agent alone.
Nothing Failed for Two Months.
Two months inside a shared repository, and no security control ever fired.
You Can't Run It. You Don't Need To.
Three different things get called having your own AI, and none of them is about which model you pick.
Fifteen Days
An accusation that fails a subtraction problem, and a breach both companies admit.
Everyone Is Calling Chatbots Agents
101 enterprises, 40% on Claude, and most "agents" are not agents.
The Off-Switch Has an Owner
Launched, switched off, switched on, metered. One model, five weeks.
"The Most Capable Model"
The most capable model is always the one that isn't out yet.
The Invisible Guardrail
Anthropic apologized for the silence, not the safeguard.
The Measurement Gap
95% of enterprise AI pilots show no return. The reason is not the model.
Claude Opus 4.8
The first AI news report produced by the news it reports on.