Your Agents Need a Save Button
summary
Argues that adding a save button — durable checkpoints of code state, files, tool calls and artifacts — lets teams freeze agents, drop compute to zero, then replay from any point with changes (model, tool, prompt) and diff outcomes. Shows Kitaru + ZenML runtime doing checkpoint replay and side‑by‑side diffs, runs cohort replays, cites DoorDash cutting what used to take hours to five minutes with 90% fewer hallucinations, and warns one replay is anecdote so checkpoint replay diff decide at scale matters.
AI-written summary · about
chapters (7)
- 0:00Agents need save button
- 1:26Replay and what-if scenarios
- 4:55Demo: Kitaru checkpointing
- 10:55Scaling replays across cohort
- 13:39Naive model swaps pitfalls
- 15:23Production replay methodology
- 16:25Conclusion
related talks
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
17mTwo Roads to Durable Agents: Replay vs. Snapshot
30mArchitecting Integrity: Responsible Software Evolution in the AI Era
10mYour agent is blindfolded
27mBreaking the Chain: Agent Continuations for Resumable AI Workflows
24mIt's Time To Rethink Everything
25mThe Ai-Assisted Developer Workflow: Build Faster and Smarter Today
27mThe AI Velocity Trap: Shipping Faster Without Breaking More