The moment your agent needs to run for more than a few minutes, everything changes. The process might crash halfway through. The context window fills up after a few hundred tool calls. The user closes their browser and comes back later expecting results. Your cloud function times out. The LLM API bill spirals because the agent got stuck in a loop at 3 AM with nobody watching.

These are not edge cases. They are the normal operating conditions for any agent doing meaningful, long-running work.

This lesson is about building agents that survive the real world. Agents that can crash and resume. Agents that manage their context window across thousands of iterations. Agents that report progress, respect budgets, and handle the async nature of tasks that outlive a single HTTP request.