I've been thinking about using the output_file parameter for a Task that will write the results of the step to a .md file (or whatever extension is appropriate), then checking for that file when running a new flow, so I can load those results instead of calling the LLM again.
What do people think of this idea?