Descriptions:
Marc Brooker, VP and Distinguished Engineer at AWS, used his AI Dev 26 x SF keynote to argue that the single biggest lever for expanding the real-world opportunity of agentic AI is not pushing the model frontier further — it is reducing the defect rate of agent outputs. Speaking from 30 years of software development experience, including daily production coding, Brooker laid out a four-quadrant framework mapping defect frequency against defect severity, and showed why only the low-frequency, low-consequence corner unlocks mass-market adoption.
Brooker observed that while the past 18 months have brought substantial progress in reducing how often agents produce errors, progress on preventing high-consequence failures in complex tasks — particularly in distributed systems and concurrent reasoning — has been much slower. He used a live anecdote about a frontier model repeatedly misrepresenting a Cauchy distribution as a fitting illustration of the problem.
He then detailed two AWS investments aimed at correct-by-construction agent outputs. The first is Hydro, a Rust framework that gives agents and human developers guardrails for building correct distributed systems and protocols — explicitly addressing current models’ known weaknesses with concurrency and failure reasoning. The second is Cedar, a policy language with deep roots in automated reasoning, designed to make authorizers provably correct. Together, Brooker positioned these as examples of a broader AWS strategy: pairing neural model improvements with symbolic and structural techniques to push agent reliability into territory where non-expert users can safely rely on the outputs.
📺 Source: DeepLearningAI · Published May 19, 2026
🏷️ Format: Deep Dive
