What Anthropic Actually Said About AI Building Itself

In June 2026, Anthropic released a report about something they call "recursive self-improvement." The title is "When AI builds itself." Most people who read the headlines think this means: AI is becoming superintelligent and will soon take over AI development entirely, building better versions of itself in a feedback loop. The actual report is asking something narrower and more honest: Can AI agents eventually pick their own research problems, not just execute them? If that happens, what does it mean? Anthropic is admitting that AI agents are getting faster at executing research. They're admitting they don't know if these agents will eventually figure out how to pick their own research problems. And they're saying that if the current trend continues, institutions aren't ready for what comes next. This article breaks down what Anthropic actually tested, what they didn't test, and what they're honestly uncertain about. It matters because if you work in AI research or engineering, th