What’s currently missing in AI agent frameworks?

Examples:

* better memory systems * workflow debugging * human-in-loop controls * distributed execution * lower latency orchestration

Interested in hearing what developers actually need.