How do you evaluate an agent when the "correct path" might involve 5 different tool calls in any order?

How do you implement "Long-Term Memory" for an agent so it remembers a user's preference from a conversation that happened three weeks ago?

What is the fundamental difference between a "Chain" (hardcoded steps) and an "Agent" (model-decided steps)? When is a Chain actually better than an Agent?

If a tool returns a massive 50MB JSON file, you can't feed that to the LLM. How do you "summarize" or "filter" tool observations for the agent?