Vol. VIII · Production Systems · VIII.1
Durable Execution
Agents run longer than request timeouts. Checkpoint after every turn. Resume as a first-class API. LangGraph won production on this axis.
6 min read
Doctrine
- Every loop iteration is a transaction: persist context, then effect, then persist observation.
- Idempotent tools, or exactly-once wrappers. Double refunds are not a prompt bug; they are a systems bug.
- Time-travel is a debugger. Keep the event log.
The production differentiator
Comparisons in 2026 keep repeating: LangGraph for durable, stateful agents. Mastra for the TypeScript equivalent. n8n for the visual DAG. Factor 6 is the doctrine independent of vendor. If you cannot pause a refund mid-approval and resume tomorrow, you are not in production.
Anti-patterns
- In-memory Python processes as the production runtime.
- Fire-and-forget tool side effects with no journal.