SANS Institute This document specifies a Credential Broker for Agents (CB4A). a credential vaulting and brokering architecture that mediates AI agent access to API credentials. Agents never hold real long-lived credentials. Instead, they receive short-lived, narrowly scoped, auditable proxy credentials issued by a broker that separates policy decisions from credential delivery. The architecture addresses the "credential sprawl" risk inherent in agentic AI, where agents aggregating access across many services become high-value compromise targets. CB4A builds on SPIFFE/SPIRE for workload identity, uses a Policy Decision Point / Credential Delivery Point separation inspired by NIST SP 800-207, and employs DPoP (RFC 9449) for sender-constrained token binding. The specification defines three credential proxy models, a tiered approval framework, and a comprehensive threat model with ten identified threats and their mitigations.