With an extremely simple ReAct agent model (call it a watered-down langchain if you might) and very dev friendly syntax to define commands, the agent (here running on GPT-4) is able to create a new command for itself to do weather forecast, activate it and use it to answer questions, in one uninterrupted live conversation. It even redesigned and redeployed it when features are missing. A second interesting thing is that, by creating one command that does absolutely nothing, calling it THINK and telling it to be methodical in the prompt, the agent is able to exhibit babyagi like behavior. Think about this as an extremely watered down version of it if you might. Curiously, just having those instructions without the dummy command does not work well.