Reinforcement Learning for Agents — with Amazon AGI Labs’ Antje Barth

Antje Barth from Amazon's AGI Labs discusses Nova Act, a new service for building reliable AI agents. She explores how they achieve over 90% reliability using reinforcement learning in 'web gyms', the shift towards 'normcore' agents for practical automation, and the future of AI as a digital co-worker.

February 3, 2026 4 min read

Reinforcement Learning for Agents — with Amazon AGI Labs’ Antje Barth — source: https://www.youtube.com/watch?v=RUklH7dDoRQ

Antje Barth, a member of the technical staff at Amazon’s AGI Labs, introduces Nova Act, a service designed to build reliable, scalable UI automation agents. The core motivation behind Nova Act is to solve the critical reliability problem in AI agents. While many agent demos appear impressive, they often only work about 60% of the time in practice. For real-world production use cases, this level of unreliability is unacceptable; as Barth states, “an agent that’s reliable 60% of the time for nearly all real-world use cases is 0% useful.” Nova Act aims to achieve over 90% reliability on enterprise workflows.

The Nova Act Developer Journey

The development process is designed to be seamless, moving from rapid prototyping to production deployment:

Playground: Developers can start for free at nova.amazon.com/act without an AWS account. They provide a website URL and describe the desired actions in natural language (e.g., navigating to a specific event page and filling out an RSVP form).
Live Preview & Debugging: The playground includes an embedded browser, allowing developers to observe the agent’s actions in real-time. It also displays reasoning traces, which are crucial for debugging and understanding the agent’s decision-making process.
Code Generation: Once the agent performs the task correctly, Nova Act generates a Python script that captures the entire workflow.
IDE Integration: This script can be imported into a preferred IDE using SDKs and extensions (e.g., for VS Code). The IDE extensions also feature an embedded live preview, keeping the developer in their flow without switching windows.
Deployment: From the IDE, the workflow can be connected to an AWS account and deployed to run reliably and securely at scale.

Achieving High Reliability with “Web Gyms”

Nova Act’s reliability stems from its unique training methodology, which moves beyond simple imitation learning. Since agents must predict the next action rather than the next word, they are trained using reinforcement learning in simulated environments called “web gyms.”

What are Web Gyms? These are high-fidelity replicas of common web UIs and workflows, such as form-filling pages or e-commerce checkouts.
Self-Play and Learning: The agent is given thousands of tasks to perform within these gyms. Through a process of trial and error (self-play), it learns to navigate, reason, and self-correct. This is analogous to how AI learned to master games like Chess and Go.
Generalization: This training helps the agent generalize to real-world websites, even when UI elements change. It learns to understand the intent behind an element (e.g., a button labeled “Create,” “New,” or a pen icon all mean “start a new draft”) rather than relying on brittle, rule-based selectors. This makes the automations resilient to website updates.

Evaluation, Security, and Governance

Benchmarking: Nova Act’s performance is validated on public benchmarks like Work Arena and RealBench, which test agents on their ability to solve tasks in unseen web environments.
Security: When deployed, Nova Act leverages the robust security foundation of AWS. All executions run within the customer’s AWS account, giving them full control over data and access. For sensitive inputs like passwords, control can be handed back to the user locally (e.g., via a terminal prompt) to ensure credentials are never sent over the network.
Observability: Detailed logs and reasoning traces are available for every step of every run, both in the IDE and in production on AWS. This provides full auditability and visibility, integrating with tools like CloudWatch.

Practical Applications and “Normcore Agents”

Amazon has over a thousand generative AI applications in development internally. The focus is on what the team calls “normcore agents”—agents that automate boring, tedious, but high-value tasks rather than flashy demos.

QA Testing: This is a primary use case. Companies like PGA Tour use Nova Act to ensure their website’s key features (e.g., weather widgets) are functional before major tournaments. Hertz automated testing for its core rental workflows, increasing their shipping velocity by 5x.
Secure Workflows: 1Password uses Nova Act to handle complex and secure sign-in patterns, a high-stakes task for a password manager.
Internal Tooling: Amazon Leo (formerly Project Kuiper) uses Nova Act to build QA frameworks across web, Android, and iOS, and even to create test cases directly from Figma wireframes before code is written.

The Future: From Tool to Digital Co-Worker

The vision at Amazon AGI Labs is to build useful and practical AI that evolves from a simple tool into a digital co-worker. This involves a multi-agent, collaborative environment where agents learn from humans and each other. By studying how humans learn and collaborate, researchers aim to build agents that can be taught tasks, ask for help, and operate with increasing independence over time. This shift will augment human productivity, allowing professionals to offload tedious work and focus on higher-level creative and strategic tasks.