In traditional service architectures, DNS / service mesh works — but in AI agent platforms where tools may be added/removed and clients need capability-based discovery, this pattern feels insufficient.

What approaches have companies used for dynamic capability discovery (not just network location) for LLM agent workloads? Would love to hear successful patterns or stumbling blocks from real implementations.

Context: I've built a minimal MCP control plane / registry prototype for enterprise discovery, but curious about patterns in practice.