A client request directed to a web site or other service hosted by a distributed pool of servers is processed by a client agent associated with the client. The client agent intercepts the client request and routes it to a particular one of the servers in the pool. The client agent bases its routing decision on address information regarding the individual servers of the pool and performance data regarding processing of previous client requests directed to the service. The address information may be retrieved by the client agent from a response to an initial client request directed to the service. The performance data may include response times for servicing previous client requests directed to the service, and the client agent may route the client request such that an average response time of multiple requests to the service is minimized. It may also make use of a probabilistic routing strategy which ensures that each of the servers in the pool receives at least one client request, such that accurate performance data may be maintained for all of the servers. A pricing manager associated with the server pool may be used to establish discount factors or other pricing information for each of the servers in the pool. This pricing information is communicated to the client agents and used in their routing decisions, so as to ensure that otherwise non-cooperating client agents will route their requests in a manner consistent with a target operating point established by the service provider.