The core problem: AI agents require measuring and iterating on subjective behavior - tone, decision-making, context usage. From experience, the best setup is when product managers take care of improving agent behavior, while engineers build workflows and infrastructure.

Restack's approach:

- Engineers build workflows in Python. Temporal and Kubernetes handle reliability and scalability.

- Product teams and domain experts A/B test and version control prompts and context management, without engineering required for behavioral iteration.

Technical stack (open source):

- React for frontend

- Temporal for retries and long-running workflows

- Kubernetes with horizontal pod autoscaler for agent scaling

- Context store built on Clickhouse

- Full observability and agent tracing

- MCP-compatible workflows