Two Roads to Durable Agents: Replay vs. Snapshot
summary
Separates durability into context durability (an append-only LLM log stored in a database) and execution durability (files, memory, subprocesses captured by OS snapshots), describes replacing replay journals with VM snapshot/restore on Firecracker microVMs, reports snapshots compressed to 14 megabytes compressed with sub-second save and ~100 millisecond restore, cites millions of snapshot-restores and 15,000 VM starts per minute, and offers an open-source CLI (FCRun) that actually works at those latencies.
AI-written summary · about
chapters (8)
- 0:00The challenge of AI agents
- 1:24History of backend architecture
- 3:26Workflow and replay models
- 5:07Agents vs. workflows
- 7:15Building durable agents
- 10:06Transition to stateful compute
- 11:50Snapshot technology evolution
- 13:45Optimized execution with fcrun
related talks
17mYour Agents Need a Save Button
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
02mHit Replay on Replay 2026
47mAgents and Memory Architecture
53mJava on CRaC: Superfast JVM Application Startup
14mMake Agents Remember the Right Context at the Right Time
30mArchitecting High-Density Performance Testing with Container-Based Simulation
40mHow to Move Fast Without Breaking Things