We already log user agents and similar for invalid requests so we can get a handle on who is forgetting to set query parameters and such. We don't have access logs on success because that would be too noisy. But having the ability to turn on access logs (for success responses) on the policy engine for short periods of time, or to have access to aggregated user-agent information would be useful for figuring out "where is this load coming from". Spun out of APPSRE-4181, where we were wondering why Envoy throughput seemed to increase when we increased the number of Cincinnati pods.