Why 80% Reliability Isn't Good Enough — Felipe Blanes, Amazon AGI Lab
AI Engineer — Oct 9, 2026 — talk
Why watch
Felipe Blanes from Amazon AGI Lab explains why AI agents that score well on benchmarks often fail in production, introducing the benchmark illusion and eval flywheel concepts. The talk covers the trust cliff (why 80% reliability feels unusable while 92% feels trustworthy) and four principles for keeping evals fresh through real customer signals. Essential viewing for AI engineers, ML engineers, and developers building production AI agents who need to move beyond static benchmarks.