Comment by talking_penguin

You're right. Monitoring is critical for this. You need to know exactly what changed and why.

This is actually where streaming + the actual architecture behing this becomes valuable beyond just "letting users watch." In Helix, every agent interaction is persisted with timing data (DurationMs), tool calls, error states, and even user feedback (thumbs up/down). Sessions track the full conversation history bidirectionally, so you can reconstruct exactly what the agent saw and did at each step.