88% of companies have experienced AI agent security incidents. But only 22% treat agents as 'identities.' Okta CEO Todd McKinnon was on The Verge and said something that caught my attention: AI agents shouldn’t just be tools; they should have their own identities. Like employees, they should log in, authenticate, and leave logs. Here’s the context. AI agents are becoming more common in companies, capable of accessing databases, calling APIs, and sending emails on their own. But most companies still manage agents by using the account permissions of their creators. What does this mean? If something goes wrong with the agent, you have no idea who instructed it, what it did, or when it did it. McKinnon’s logic is: agents need independent identities, independent permissions, independent logs, and even a kill switch. If an agent behaves abnormally, you should be able to shut it down with one click. I predict that agent identity will become a core topic for enterprise AI in the second half of 2026. Whoever nails this infrastructure first will become the toll booth for the next wave of AI infrastructure.