Okta announced a new set of product features on July 22 aimed at closing a gap that's been widening as enterprises deploy AI agents: securing not just an agent's initial login, but every connection it makes afterward, to tools and to other agents, for the life of that connection. The announcement describes new innovations within Okta for AI Agents built to help organizations secure agent connections to enterprise tools and other agents at runtime, while governing those connections over time. That framing matters. Most identity tooling to date has focused on the moment an agent authenticates. Okta's pitch here is about what happens after: an agent that received a scoped credential six weeks ago may still be holding it, whether or not the task it was granted for is still active. The stated goal is to prevent standing and excessive permissions from accumulating unnoticed, which is the same failure mode that has quietly plagued human IAM for years and is arguably worse for agents, since they can spin up sub-tasks and delegate to other agents without a person watching each hop. A credential that made sense for a narrow, bounded job can end up reused across a chain of delegated calls that nobody explicitly authorized. This lands the same week other vendors and standards bodies have been pushing on adjacent pieces of the same problem. The A2A protocol shipped a stable 1.0 specification earlier this year that added signed Agent Cards, mTLS, and formal JWS-based signature verification, moving agent identity from self-declared metadata toward something a receiving party can actually check. Separately, groups like the Decentralized Identity Foundation have taken up frameworks using DIDs and verifiable credentials specifically to answer who authorized an agent and what it's allowed to do on a given task. What's notable about the Okta move is the emphasis on ongoing review rather than a one-time grant. Access that was appropriate when an agent was provisioned doesn't necessarily stay appropriate as the agent's role, scope, or delegation chain evolves. Periodic certification of what an agent can still reach, rather than assuming initial authorization holds indefinitely, is a small but concrete step toward making agent permissions auditable in the way enterprise IAM already treats human accounts. It's a narrow fix, but it targets a specific, named failure mode rather than agent security in the abstract.