What is the biggest challenge in building reliable AI agents for real-world applications: planning, tool execution, memory, evaluation, or handling unexpected failures? What practical approach has worked best for you?