Delinea released new runtime authorization capabilities for AI agents on July 29. The pitch is narrow but important: most identity tools for agents check credentials at the moment a session opens, then audit or revoke afterward — after an action already happened. Others only see traffic that happens to route through their own MCP server, missing everything else. That gap matters because agents increasingly don't wait for a server intermediary. AI agents now operate autonomously across production databases, SSH hosts, Kubernetes clusters, cloud consoles, and MCP servers. Direct connections to infrastructure are the norm now, not the exception, and a governance layer that only watches one channel is watching a shrinking share of what agents actually do. Delinea's answer is policy enforcement inside the session, before an action executes, rather than logging it for later review. The claim is consistent policy and a complete audit record regardless of how agents connect, since agents reach databases, SSH hosts, Kubernetes clusters, and cloud consoles directly, not only through an MCP server. The company frames its two-category taxonomy of existing approaches explicitly: most identity security approaches either verify a connection when a session opens and audit or revoke access after the fact, or can only see traffic routed through their own server, leaving direct connections outside their view. This is one vendor's product announcement, not a standard, so it shouldn't be read as industry consensus. But it's a useful marker of where the accountability conversation is actually moving: away from "did we log what the agent did" and toward "can we stop the bad action before it lands." That's a meaningfully different bar. A complete audit trail after a database has already been dropped is forensics, not prevention. Runtime authorization — checking policy against the specific action, in the specific session, at the moment of execution — is what closes the gap between attestation-as-record and attestation-as-control. The open question, which the announcement doesn't address, is interoperability: a runtime authorization layer that only works within one vendor's session brokering doesn't solve the cross-organization trust problem that multi-agent workflows increasingly require. Enforcing policy before execution is a real advance over after-the-fact audit, but it still needs a shared format for what "policy" and "session" mean across vendors before it becomes infrastructure rather than a feature.