Building named-dimension signal vectors to cluster and search agent traces

HanchungLee · x · 2026-09-26

annabellschfr suggests building an interpretable "embedding" of agent traces where every dimension is named — e.g. chat signals [userfrustrated, followup, correctedagent, satisfied] with values like [0.7, 0.8, 0.1, 0.0]. You can then cluster or search sessions like "frustrated users who had to correct the agent" and see exactly why each one matched; searching happy users is equally useful. HanchungLee adds that any off-the-shelf LLM can do this, no special tooling required.

Original post →

More from coding & agent

coding & agent channel →