Designing the rulebook for your company's AI agents

AI agents were joining company workforces faster than anyone had figured out how to manage them. A third-party agent could be built on Google, hosted on AWS, and deployed inside Microsoft 365, and nobody had a single place to see it, secure it, or shut it down if something went wrong.

The problem wasn't theoretical. As enterprises started deploying agents at scale, the gaps became real, fragmented identity systems, no visibility into what agents were doing, no consistent way to set permissions or respond to incidents. Managing an AI workforce with the tools built for human workforces wasn't working.

Project Kairo was Microsoft's answer. A small team was pulled together to envision what a unified agent governance platform could look like, spanning security, identity, performance, and management across every agent in an organization regardless of where it was built or hosted.

Four of us were brought in for what felt more like a startup sprint than a typical Microsoft project. We had a demo script broken into narrative beats, a rough sense of the problem space, and a couple of weeks to make it feel real.

Each designer owned a distinct area. I focused on security and incident management, designing the screens that would show administrators what their agents were doing, what went wrong, and what to do about it. We'd mock up screens independently, present to each other for critique, refine based on feedback, then do the same with product leads before packaging everything for senior leadership.

The closest references we had were HR tools like Workday and IT configuration management platforms like ServiceNow, but for AI agents instead of people. We also looked closely at existing Microsoft products like Defender and Entra to make sure the patterns felt native to the ecosystem rather than something bolted on. We weren't copying those patterns, we were asking what they'd look like if you rebuilt them from scratch for a workforce that never sleeps, operates across clouds, and can act autonomously on behalf of users.

The security and incident management screens were designed around a simple question: what does an administrator actually need when an agent does something it shouldn't?

The answer we landed on was immediacy and clarity. A push notification to your phone at 2AM telling you a third-party agent just attempted to create an unauthorized admin account. A full incident view showing exactly what the agent tried to do, what policy it violated, the evidence trail, and a single button to quarantine it. An activity log that reads like a conversation, showing the chain of agent actions that led to the violation in real time.

The Agent Management Center overview was a collaborative effort across the team. The agent landscape visualization, showing every agent across every department mapped as an interconnected network was something none of us had seen before because it didn't exist before. Neither did platform health rankings showing which third-party agent platforms were generating the most incidents, or AI-generated teaching opportunities surfacing patterns across the agent workforce.

For the executive presentation I built dark mode versions of the key screens. Sometimes you need a room to feel something before they fund it.

Project Kairo got the greenlight. The envisioning work was presented to senior Microsoft leadership and fully funded as a new product and business line. It was announced at Microsoft Ignite in November 2025 as Agent 365 and went generally available on May 1, 2026.

Four designers. Two weeks. No existing reference point. A product that ships to enterprise customers worldwide.