Sentient team published the most heretical agents paper of the year. Microsoft, Alibaba, and Google all cite EvoSkill: Automated Skill Discovery for Multi-Agent Systems. Sentient researchers and Virginia Tech co-authors open by showing a Claude agent confidently botching a Treasury question and only get more unhinged from there. They demand the AI industry abandon its biggest obsession, hand-crafted skills. Right now, everyone from solo devs to AI labs assumes hand-crafting is the only path. An expert studies the domain, then writes the playbook for the machine. Sentient argues that this entire workflow is a manual illusion. Agents do not need "human" teachers. They are flexible general-purpose machines, held back simply by the domain expertise we never gave them. We only think skills must be written by hand because we are completely blind to the millions of lessons hiding inside the failures we never show them. Which brings us to the Treasury argument. Claude with Opus 4.5 is one of the greatest coding agents in history. But dropped into 89,000 pages of Treasury data, where humans average 50 minutes per question? It is fundamentally lost. Our belief that frontier agents are "ready" out of the box is pure benchmark bias. They aren't objectively prepared. They're just better than the rest, which are hopelessly worse at it. Sentient says we need to stop writing playbooks for our machines. Instead, they propose a new loop: EvoSkill. Automated Skill Discovery. Instead of trying to author every skill with our flawed, time-limited human hands, we need to embrace self-evolution. EvoSkill is about the speed of learning from failure. It is a system where one agent fails a task, a second autopsies the failure trace, and a third writes a brand new skill to patch the gap. The model stays completely frozen. More importantly, the skills it discovers are designed to transfer where hand-crafted ones never could. Things like exploding 12 points on noisy web search. Or jumping 7 points on Treasury reasoning. Or transferring zero-shot to a benchmark it never saw, and still gaining 5. The entire AI industry is obsessed with humans writing playbooks in our own image. Sentient's paper is a brutal wake-up call.