“The finite context length of LLMs is a key limitation for agent systems, restricting the inclusion of historical information and detailed instructions.”