According to one embodiment, a method, computer system, and computer program product for software agent synthesis is provided. The present invention may include generating one or more training examples from historical data and software agents; training, using the on the one or more training examples, a language model to synthesize a software agent based on a natural language input from a user; monitoring, using one or more input devices, for one or more natural language user inputs; responsive to identifying one or more natural language user inputs, synthesizing, using the trained language model, one or more software agents based on the one or more natural language user inputs; execute the one or more new software agents to carry out one or more tasks invoked by the one or more natural language user inputs.