How Do You Judge the Execution Quality of a Sub-Agent in Multi-Agent Orchestration?

How Do You Judge the Execution Quality of a Sub-Agent in Multi-Agent Orchestration?

In newtype Profile, my approach is to give each specialized Agent multi-dimensional quality scores. For example, Researcher is an intelligence officer, so its three evaluation dimensions are coverage, source quality, and relevance; Writer is a writer, so its three evaluation dimensions are structure, clarity, and being well-supported.

When the score is above 0.8, it can be handed off to the next Agent. Between 0.5 and 0.79, it should optimize itself. If it is below 0.5, it has to be redone.

The key point of this design is that the system precisely points out the weakest dimension and gives targeted improvement suggestions, rather than vaguely saying, “Try again.”