TL;DR — Key Takeaways
- Classie Supervise continuously discovers and monitors sanctioned and unsanctioned AI agents operating across endpoints, browsers and enterprise IT environments.
- The platform uses Open Policy Agent (OPA) rules and lightweight inline sensors to enforce controls in real time while creating transcripts of agent activity.
- As agents act at machine speed, Classie is positioning real-time intervention, auditability and policy enforcement as essential for limiting the blast radius of rogue or unexpected behavior.
Classie today made available a platform that makes it possible for organizations to both continuously monitor sanctioned and unsanctioned agents and enforce policies in real time.
The Classie Supervise platform defines rules via an Open Policy Agent (OPA) policy engine that are then enforced inline by lightweight sensors. That capability makes it possible to then create runtime transcripts of agentic workflows that surface who initiated the activity, what the agent did, what information it used, the total cost, along with an audit trail.
Collectively, those capabilities make it possible for humans to not only redirect or terminate an agentic AI action but also prevent rogue behavior, says Classie CEO Poonacha Kongetira.
Designed to be installed in 15 minutes, Classie Supervise typically requires about five days to discover and monitor agentic AI workflows. It tracks agent activity across endpoints, browsers, and enterprise IT environments to create a transcript that connects agent and user identity with the context, intent, data being accessed and actions.
The overall goal is to capture the exhaust of AI agent activity to give organizations the governance capabilities needed to safely manage and deploy AI agents, says Kongetira. “It established a chain of custody for AI agent activity,” he adds.
Priced at $99 per user, the availability of Classie Supervise comes in the wake of a series of high-profile incidents involving rogue AI agents. While those incidents have sparked more calls for regulations, the fact remains that many organizations have already deployed multiple AI agents. The challenge now is making sure they are able to maintain control of AI agents that have been programmed to complete tasks by any means necessary despite what guardrails might have been embedded in an instruction.
Each organization will need to determine its own level of comfort with deploying AI agents, but there are already probably more of them running than the leaders of an organization may realize. Many employees are configuring and deploying AI agents to automate various tasks without always asking permission. As such, the number of unsanctioned AI agents might already far exceed the number of ones that have been officially sanctioned.
In some cases, a centralized IT team may be able to discover and remove those AI agents. However, it’s more likely that the productivity gains provided by AI agents justify the risk, so it then falls to IT and cybersecurity teams to minimize the potential blast radius of an incident.
Regardless of approach, most organizations are going to experience multiple incidents involving both sanctioned and unsanctioned AI agents. The challenge and the opportunity are to prevent as many of those issues from occurring in the first place and, if needed, to be able to apply controls in real time to keep the amount of potential damage wrought to an absolute minimum. The issue, of course, is that AI agents are performing tasks at machine speed, so in the absence of controls that can be applied in real time as aberrant behavior is detected, the size of the blast radius of an incident involving AI agents is only going to increase exponentially as every second ticks by.
