Zero-Knowledge Credential Injection: What are the latency and scaling trade-offs?

The idea of AI agents never directly holding credentials, instead relying on OS keychains and local proxies for zero-knowledge injection, sounds incredibly secure. From an architectural standpoint, I'm curious about the practical implications for latency and scaling. For agents needing frequent access to multiple credentials, could this 'proxying' introduce noticeable overhead? And how well would such an infrastructure scale for thousands or millions of agents making simultaneous authenticated calls? Are there specific types of agent workloads where this approach might be less performant, or conversely, areas where the security benefits so overwhelmingly outweigh any potential performance hit that it's a clear win?