Phase 4 — Production & Depth · Lesson 42 · 20 XP
Durable workflows for long-running agents
An agent task that runs for minutes or hours can't safely keep all its state only in one process's memory — a crash, a redeploy, or a timeout loses everything and forces a full restart from scratch.
Durable execution checkpoints state after each step, so a workflow can pause, resume from exactly where it left off, or survive a crash mid-run. Tools built for this (Temporal, the Vercel Workflow DevKit) structure work as explicit steps with automatic retries per step, rather than one long unbroken function.
Exercise
Take a multi-step agent task and break it into explicit steps with checkpointed state persisted after each one — so it could genuinely resume from the last completed step instead of restarting from the beginning after a crash.
Check yourself
1. Why can't a long-running agent safely keep all of its state only in memory?
2. What does "resume from the last completed step" actually require your code to persist?
Reliability: retries, fallbacks, queues
Answer the check-yourself questions to unlock this