Agents need watchdogs, and now Windows ships one
Microsoft's Execution Containers and Anthropic's oversight rule point the same way: agents get a leash.
For a year the pitch was 'let the agent do it.' The quiet correction in late October 2026 was 'let the agent do it, inside a fence.' Two moves said it loudly.
Microsoft made Execution Containers generally available on Windows 11. In plain terms, an agent can be given a sandbox that caps what files and networks it touches. That is the difference between an assistant that might email your boss and one that cannot reach your boss.
Anthropic set a usage policy, effective November 12, requiring human oversight for equipment that could cause injury and expanding verified access for authorized security work without dropping safeguards. The message is that capability without oversight is not a feature; it is a liability.
I am skeptical of most agent hype, but this I like. A watchdog is cheaper than an incident, and a sandbox is the rare security control that helps the user instead of nagging them. The agent still runs; it just cannot quietly exceed its brief.
The open question is enforcement. A container only works if the app respects it, and a policy only matters if it is checked. But the direction is right, and it is the most grown-up thing to happen to agents this year.
Frequently asked questions:
Q: Do Execution Containers stop all agent mistakes?
A: No, but they cap blast radius. A mistake stays in the sandbox instead of your inbox.
Q: Who does the Anthropic oversight rule affect?
A: Anyone building agents near physical or security risk. Human-in-the-loop becomes required, not optional.
Q: Is this just enterprise theater?
A: The container part is practical today. The policy part only works if someone actually checks.