Two Roads to Durable Agents: Replay vs. Snapshot — Eric Allam, Trigger.dev

Replay-based durability — wrapping every step in a journal, replaying on recovery, requiring deterministic code — is how everyone makes agents durable today. It works until it doesn't: the journal grows with every turn, the structure starts constraining how you write code, and an agent that needs to run for hours starts looking less like a transaction and more like a session. This talk separates the problem in two: context durability (the append-only log of everything the LLM saw, which already fits in a database) and execution durability (the files, memory, and subprocesses that live in the compute layer, which don't). The answer to the second half isn't a smarter log — it's OS-level snapshot and restore. Eric Allam walks through how Trigger.dev built this on Firecracker microVMs, getting snapshots down to 14 megabytes compressed with sub-second save and hundred-millisecond restore times, and why IBM mainframes in 1966 got there first. Speaker info: - https://x.com/maverickdotdev - https://www.linkedin.com/in/eric-allam/ - https://github.com/ericallam