Getting an agent to finish a task is easy. Getting it to finish after a crash, a timeout or a three-day wait for approval is something entirely different. As agents take on longer, multi-step work that writes to real systems, a single failure halfway through can leave half-finished actions, duplicated calls and state nobody can reconstruct. So how do you build agents that can stop, resume and undo without starting over? This session looks at how builders are applying durable execution to agent runtimes, from checkpointing agent state and replaying from recorded history to making every side-effecting step idempotent. Through real architecture patterns and failure cases, it covers what it takes to make long-running agents recoverable, and how to roll back work an agent has already done when something goes wrong. Key takeaways: → How checkpointing and replay keep long-running agents from restarting from zero → Designing idempotent tool calls so retries never duplicate real-world actions → Using compensating actions to unwind partial work after a failure