LLMs, GenAI, RAG, agents, evaluation · insight-1565

Run a new write-capable agent in shadow mode and compare its proposals with human actions first

Situation
A team wants to give a drafting agent permission to apply refunds or rebookings after good offline test results.
Action
Let the agent run on live cases with its write tools replaced by recorders, compare each proposed action with what staff actually did, review disagreements by class, and grant write permission per action class once agreement and error cost meet a threshold.
Reason
Shadow runs use real inputs without real effects, so they expose gaps that curated tests miss before any harm occurs.
Boundary
Agents that will gain write access to real systems.

Symptom or risk

Real traffic contains cases the test set lacks, so the first live writes reveal the gaps as customer harm.

Cases where the human action is itself unreliable or unobserved give a weak reference.

Weeks of parallel running and review effort; staff actions are an imperfect label.

Counting agreement only and ignoring which disagreements are costly.