Goal-oriented dialog systems require a different approach from chit-chat conversation systems in that they typically interact with an external knowledge base. It is desirable to incorporate domain knowledge to operate on the external knowledge. This paper presents extensions to hybrid code networks with reinforcement learning agent for the sixth dialog system technology (DSTC6) Facebook AI research (FAIR) dialog dataset. It reduced the effort in manually constructing the rules or additional labeling compared to the previous approaches. Thanks to the well-designed RL agents and reasonable domain-specific rules, the proposed model achieved high accuracy in the most of the test sets.