I agree about the need for an environment. The difference between GPT-3 and an agent in an environment is that GPT-3 only saw tons of static text, while an agent can design an action (an experiment), act it out and observe the results, drawing conclusions. Thus it can act in a way similar to the scientific method.