The Cloud Security Alliance has published a purpose-built identity and access management framework for agentic AI, and the framing is notable: it doesn't try to bolt agents onto human IAM. It starts from the premise that agents break the model traditional IAM was built for. The publication introduces a framework that accounts for autonomy, ephemerality, and delegation patterns of AI agents in complex multi-agent systems. Where human IAM assumes a relatively stable subject — one person, one set of credentials, occasional role changes — agents spin up, delegate authority to sub-agents, and disappear within a single task. That churn is the actual problem, not a variant of an old one. The blueprint leans on decentralized identifiers and verifiable credentials, paired with Zero Trust principles, to manage agent identities, and it addresses secure delegation, policy enforcement, and real-time monitoring as first-class concerns rather than afterthoughts. It also walks the full lifecycle: identity creation and attestation, runtime authorization, logging, monitoring, and incident response — treating an agent's identity as something that has to be verified and re-verified continuously, not issued once and trusted forever. The framework leans on CSA's existing MAESTRO threat-modeling model to analyze security considerations, and it compares centralized, decentralized, and federated deployment options rather than prescribing one architecture for every organization. That's a reasonable hedge: a bank running internal agents on a controlled network has different delegation risks than a marketplace where agents from different vendors negotiate with each other. What makes this worth flagging isn't novelty of the individual pieces — DIDs, VCs, and Zero Trust have all been discussed for agents before — but that a major industry body is now packaging them as a coherent lifecycle spec rather than scattered recommendations. Enterprises adopting agents at scale have been improvising IAM policy agent by agent. A reference architecture, even an early one, gives security teams something concrete to measure their own gaps against, and gives vendors a shared vocabulary for what 'agent identity' is actually supposed to mean when they claim to support it. The gap this doesn't close: guidance is not enforcement. Whether DIDs and VCs for agents get implemented consistently, or become another checkbox in a compliance PDF nobody operationalizes, depends on tooling and adoption pressure that hasn't shown up yet.