A product lead at Arise shares lessons learned building an AI agent (Alex) over nearly a year, focusing on context management strategies. The talk covers a vicious loop of context overflow failures, why naive truncation and LLM summarization both failed, and how a 'smart truncation with memory store' approach (keeping head and tail of context, storing the middle for retrieval) became their working solution. Key insights include: context engineering matters more than prompt engineering for agent success, long sessions require dedicated evaluation (loading 10 turns and testing the 11th), and offloading heavy data operations to sub-agents keeps the main conversation context lean. Open challenges include long-term memory across sessions, principled context budgeting, and handling ever-growing conversation lengths.