Can AI Agents Police Other AI Agents? Microsoft's Guardian Agents Raise Fresh Questions

Microsoft has introduced Project Perception, which adds what it calls Guardian Agents. These are designed to monitor AI agents and contain any that behave in an anomalous way. The effort is part of Microsoft's broader bet on a unified security stack, where monitoring and response sit within a single set of tools.

Not everyone is convinced. Experts quoted in the coverage warn that using one set of unpredictable agents to police another may multiply the risk instead of containing it. The concern is that the oversight layer is itself made of the same kind of technology it is meant to control.

The framing in the article also matters: AI agents are being treated as a form of insider threat. Like a trusted employee with broad access, an agent that acts unexpectedly can cause harm from inside the business. For small businesses adopting AI tools, this is a useful reminder that these agents need the same careful thought about access and oversight as any staff member.