https://video.ut0pia.org/videos/watch/c1fe0019-9306-4a27-a3b8-33b7cad42e7b 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 Mon, 11 May 2026 17:36:01 GMT https://validator.w3.org/feed/docs/rss2.html PeerTube - https://video.ut0pia.org https://video.ut0pia.org/lazy-static/avatars/0287a09a-aae7-4840-9843-b416426e7046.webp https://video.ut0pia.org/videos/watch/c1fe0019-9306-4a27-a3b8-33b7cad42e7b All rights reserved, unless otherwise specified in the terms specified at https://video.ut0pia.org/about and potential licenses granted by each content's rightholder.