One thing I've done is have some tool calls not return the result directly, but rather write them to a shared state, where they are put into the system context on the next turn. I did this because the LLM kept calling the `read_file` tool and blew up it's context. This "shared cache" fixes that by only allowing one copy of a file to ever be in the context/prompt. There was also emergent behavior of them using the cache for storing plans or thoughts, so I made that a first class principle and it "works" pretty great (making, following, updating the plan), but it is unclear if it actually results in better outcomes (insufficient evaluation capabilities, generally in the field)
I'm still unsure if the file-in-cache idea is a good one, but at least I can now find out with the deep access and customization of my agentic setup
One piece of advice I received just before starting this journey was...
"Give your agents great tools and then get out of the way"