Observability, error handling, and reliability in workflows

What features, tools, or approaches people consider essential for making agent workflows robust. For example: retry policies, failure detection, logging, tracing, checkpoints. How does GraphBit align with what folks need?