Claude can now work unsupervised for days. Your move, interns.
What happened
Anthropic released an agent mode that lets Claude run long-horizon tasks on a computer for days at a time, checkpointing its own progress.
Why this matters
The unit of AI work is shifting from "answers" to "outcomes". Pricing, benchmarks and jobs all get rethought around that.
The question is no longer "can it code?" but "can you trust it with the keys?"
The slightly cynical read
Every demo shows the agent succeeding. Nobody shows the Slack message asking why prod is down.
What to watch next
Whether enterprises actually grant these agents real permissions, or keep them in sandboxes forever.