Why 80% Reliability Isn't Good Enough — Felipe Blanes, Amazon AGI Lab
Your agent scores great on benchmarks. Then real customers use it and it breaks.
Felipe Blanes from the Amazon AGI Lab shares what he learned working directly with customers of Nova Act, Amazon's service for building browser agents, from research preview to general availability on AWS. He calls the core problem the benchmark illusion: static evals look great until customers do things nobody expected, and then all you have left is hope. His fix is an eval flywheel built around the customer: define success the way the customer does, capture signals (including by talking to customers), diagnose gaps in the model, harness or product, and feed them into decisions. He explains the trust cliff (80% reliability feels like more work; around 92% feels trustworthy), why being open about limits builds more trust than higher benchmarks, surprising use cases from Amazon Leo, Hertz and Sola, and four principles for keeping evals fresh.
The text is the source's own description of its publication. The content belongs to AI Engineer.
More from AI Engineer
Inside Arca's Long-Horizon Agents That Run the Business - Aman Singh, Arca
AI Engineer 23 Oct
Real AI Agent Attacks and What Would Have Stopped Them — Andrey Rybka, Bloomberg
AI Engineer 23 Oct
Outside of tech there has been no financial payoff to US companies for using AI, now viewed as a margin-hurting cost center, as VC-backed firms burn more tokens
Techmeme 4 h ago
Source: Replit recently moved most AI spend from Anthropic to OpenAI, which gave 30%+ off in a deal worth $100M+/year, after Anthropic discontinued its discount
Techmeme 6 h ago