Eric Allam (Trigger.dev co-founder) explores two approaches to making AI agents durable in production: the replay model (caching side effects as steps in a journal) and the snapshot/restore model. He argues that replay, which underpins 30 years of stateless backend infrastructure, breaks down for long-running agents because the journal grows unboundedly. Instead, durable agents need two things: a persistent append-only context log (all LLM messages and tool calls) and execution-layer snapshots. Trigger.dev evolved from CRIU-based process checkpointing to Firecracker microVMs, achieving sub-second snapshots and ~200ms restores, compressing 512MB VM state down to ~14MB. They are open-sourcing this as a Docker-like CLI tool called FCRun, capable of 15,000 VM starts per minute.