DeepSeek just told everyone how the sausage gets made, and the honest answer is: lots and lots of sandboxes.
The company published a paper describing the platform behind its AI agent training, which runs about 3 million sandboxes daily. Each sandbox is an isolated environment where an agent can act - clicking, coding, executing commands - without touching anything real. The paper's more interesting admission is that agent execution is untrustworthy by default. DeepSeek states plainly that no single safeguard can catch every way an agent might misbehave during training.
That's a notable thing for a major AI lab to put in writing. Most agent announcements lead with capability numbers, not with "our system might do something wrong and we can't fully stop it." It suggests the real bottleneck in scaling AI agents isn't model quality anymore - it's containment infrastructure and the layered defenses needed to run millions of untrusted, semi-autonomous processes a day without something going sideways.
Regulators are chasing a version of this same problem with a very different tool: Europe requires each member state to stand up its own regulatory sandbox. One is a testing ground for compliance, the other for containing bad agent behavior at industrial scale - and DeepSeek is running its version 3 million times a day before breakfast.