Antje Barth

Antje Barth is a Member of Technical Staff at Amazon AGI, where she works on making advanced AI systems practical for developers and reliable enough for real-world tasks. Her career spans enterprise infrastructure, machine-learning education, generative AI, and agents that can understand interfaces, collaborate with people, and verify their own work.

From cloud infrastructure to autonomous systems

Before joining AWS, Barth worked in solutions engineering and technical evangelism at Cisco and MapR, with a focus on infrastructure, large-scale data systems, containers, and Kubernetes. She later became a principal developer advocate for generative AI at AWS and co-founded the Düsseldorf chapter of Women in Big Data and the Generative AI on AWS Meetup.

With Chris Fregly, she wrote Data Science on AWS, published in 2021, covering the practical lifecycle of machine-learning systems from data preparation through deployment and operations. In 2023, Barth, Fregly, and Shelbee Eigenbrode published Generative AI on AWS, addressing model evaluation, fine-tuning, retrieval-augmented generation, and production deployment. Its open-source companion repository supplies working examples.

Barth also co-instructed Generative AI with Large Language Models, developed with DeepLearning.AI and AWS. Her teaching frames AI engineering as an accessible discipline that nevertheless requires serious attention to model selection, privacy, security, evaluation, and responsible application design.

• Retrieval-based tool selection: Agents with thousands of available tools cannot fit every description into a model’s context. Barth has described storing tool descriptions in a knowledge base and retrieving only those relevant to the immediate task. Her cloud-scale agent architecture also incorporates Model Context Protocol servers, serverless deployment, session management, and explicit authorization.

• Perception agents: Reliable computer-use agents must understand rendered interfaces, detect changing application state, and assess whether their actions succeeded. Barth applies a perception-planning-action loop to software environments, allowing agents to work across applications even when dedicated APIs are unavailable.

• Shared visual context: Browser-based annotation lets someone identify the precise interface element they want changed, giving an agent concrete visual and structural context instead of an ambiguous textual description.

• Visual verification: Coding agents can rely on executable tests; most knowledge work offers no equivalent safeguard. Barth’s open-source perception-agent approach checks rendered output against design specifications, exercises application flows, and reports failures for human review. She has also demonstrated how spoken design discussions can trigger interface changes followed by the same verification process.