A developer prototyped an agent by looping over Messages API calls and executing tool calls by hand. The loop now needs context compaction, permission prompts for risky tools, and session resumption. What is the most direct way to add these?