Agentic Systemsadvanced13 min · 9 steps
Durable Execution
Keep long-running agents alive through crashes and deploys: journal every step, replay deterministically, make side effects exactly-once, and pause for days without burning compute.
errordatarequesttool callresponsesignal
ProblemStep 1 of 9
Crash after the charge, before the record
The agent calls a tool that charges a card. The charge commits — then the worker is killed by a deploy before it records the result. On restart, a naive retry runs the step again: two charges.
After a mid-step crash you can't tell whether the side effect committed or only its acknowledgement was lost. A checkpoint saved between steps still leaves that unsafe window.
Watch outCheckpoints capture state. Durable execution captures the journal of steps — and you can't safely resume mid-side-effect from a snapshot alone.
Lesson 11 of 15 in Agentic Systems