Solid roundup. The thing I'd add as the gateway becomes load-bearing for agents specifically: it stops being just rate-limit plus auth and becomes the model-routing and cost-control layer. For agentic workflows the killer feature is per-call routing (cheap model by default, escalate to the expensive one only when the task needs it) plus hard spend caps so a runaway loop can't quietly bill you a fortune overnight. Plain HTTP gateways don't think about token cost or which model; an AI gateway has to. That routing-plus-budget-guard layer is exactly what I built into Moonshift so an agent run stays a few dollars instead of spiraling. Of the ones you listed, which actually do per-request model routing versus just proxying the call?