Hi everyone,
I’ve put together a reference implementation for building fault-tolerant, long-running multi-agent AI orchestrations with Temporal and CrewAI, including explicit Human-in-the-Loop (HITL) approval gates.
To keep workflow execution deterministic, I isolated CrewAI’s non-deterministic imports from the workflow sandbox. State is managed through explicit Signals, and workflow.patched() handles versioning without downtime.
I’ve documented the core design patterns and open-sourced the repository. Feedback on the approach is welcome.
Read the full breakdown on dev.to:
Explore the codebase on GitHub: