WorkAny

最新的界面做好了！如果大家喜欢的话，这个版本的skill我们也开源😆

Something strange is happening with AI agents that this new Anthropic research quietly surfaces. The agents are asking us for help more than we're stepping in to correct *them*. Anthropic analyzed data from Claude Code and their public API to measure how autonomous AI agents actually are in practice. The headline finding is what you'd expect. Agents are running longer (sessions nearly doubled, from 25 to 45 minutes), and experienced users approve more actions automatically. All that's great! But here's the part worth sitting with. Claude Code stops to ask for clarification more than twice as often as humans manually intervene. The agent is, in a real sense, more cautious about its own limits than its operators are. That's a weird inversion of the usual AI safety framing, where we assume the human is the responsible one pumping the brakes. The practical safety numbers are reassuring on the surface. 80% of tool calls have at least one safeguard; only 0.8% of actions are irreversible. But I think the more important finding is buried in the recommendation. They argue that effective oversight "doesn't require approving every action but being in a position to intervene when it matters." That's a subtle but significant shift. The research argues that experienced users stay vigilant even with auto-approve on. Maybe. But as sessions get longer and approval becomes muscle memory, the gap between being in a position to intervene and not paying attention starts to collapse.