I've found this article pertinent: "Anti-patterns while working with LLMs".

These key points are useful - likely even if you're implementing Agentforce:

👉 Repeating the same context wastes tokens, so only send updated information instead of duplicate data.

👉 Don't give tasks LLMs where they perform poorly - generate code for precision tasks instead.

👉 If you let context windows get too long you will lose accuracy and distort memory. Manage the context actively.

👉 Models are better at widely discussed topics than niche or obscure domains.

👉 Delegating too much judgment to an LLM leads to unnoticed errors. Subtle mistakes can be harmful.