Super Data Science: ML & AI Podcast with Jon Krohn 963: Reinforcement Learning for Agents, with Amazon AGI Labs’ Antje Barth
78 snips
Feb 3, 2026 
Antje Barth, technical staff and developer relations lead at Amazon AGI Labs and bestselling O’Reilly author, discusses Nova Act and building reliable AI agents. She describes the playground for rapid prototyping, RL web gyms that train agents on thousands of UI tasks, benchmarks and observability for production reliability, and security and AWS integrations for safe deployments.
AI Snips
Chapters
Transcript
Episode notes
Train Agents In Web Gyms
- Train agents via reinforcement-learning web gyms where they self-play thousands of UI tasks to learn robust actions.
- This trial-and-error training improves generalization when UIs change (buttons, icons, layouts).
Evaluate With Benchmarks And Real Use Cases
- Evaluate agents on benchmarks like RealBench and on your company's real tasks for meaningful metrics.
- Work closely with customers to run evaluations on the specific enterprise workflows you need to support.
Keep Control By Running In Your AWS Account
- Use AWS integration for production: run executions in your AWS account to retain security, observability, and control.
- Leverage traces, step views, and CloudWatch to debug and audit each run before scaling.