In this paper we present an extension of Belief-Desire-Intention agents which can adapt their performance in response to changes in their environment. Specifically we consider situations in which the agent’s actions no longer perform as anticipated by the programmers who created the agent’s plans. Our agents maintain explicit descriptions of the expected behaviour of their actions, are able to track action performance, learn new action descriptions and patch affected plans at runtime. Our main contributions are the underlying theoretical mechanisms for data collection about action performance, the synthesis of new action descriptions from this data and the integration with plan reconfiguration. The mechanisms are supported by a practical implementation to validate the semantics.