We discuss a packet network architecture called a cognitive packet network (CPN), in which intelligent capabilities for routing and flow control are moved towards the packets, rather than being concentrated in the nodes and protocols. Our architecture contains "smart" and "dumb" packets, as well as acknowledgement packets. Smart CPN packets route themselves, and learn to avoid congestion and losses from their own observations about the network and from the experience of other packets. They use a reinforcement learning algorithm to route themselves based on a goal function which has been assigned to them for each connection. Dumb CPN packets of a specific quality of service (QoS) class use routes which have been selected by the smart packets (SPs) of that class. Acknowledgement (ACK) packets are generated by the destination when an SP arrives there; the ACK heads back to the source of the SP along the inverse route and is used to update mailboxes in CPN routers, as well as to provide s ource routing information for dumb packets. We first summarize the basic concepts behind CPN, and present simulations illustrating their performance for different QoS goals, and analytical results for best and worst case performance. We then describe a test-bed network we have designed and implemented in order to demonstrate these ideas. We provide measurement data on the test-bed to illustrate the capacity of the network to adapt to changes in traffic load and to failures of links. Finally, we use measurements to evaluate the impact of the ratio of smart to dumb packets on the end-to-end delay experienced by all of the packets. © 2001 Elsevier Science B.V. All rights reserved.