The Agentic Systems Lab is a significant contributor to the agentic AI stack, particularly in the areas of evaluation and orchestration. They are active in the development of 'Next-Gen RAG' frameworks like AutoRAG and MetaRAG, which are essential for building reliable, data-grounded agents. For the broader ecosystem, the lab provides the benchmarks necessary to move agents from experimental prototypes to production-grade tools by defining how safety and performance should be measured.
They are also notable for their support of the Model Context Protocol (MCP), specifically through partnerships with projects like mcp-use.com. By training the next generation of 'AI builders' and fostering startups that focus on generally capable agents, the lab is helping to define the standards for how agents interact with software interfaces and simulated environments. Their work on World Models is particularly relevant for the future of agents that need to plan across long horizons in dynamic, real-world settings.
The Agentic Systems Lab is an AI research entity at ETH Zürich focused on the intersection of multimodal models and real-world automation. Based within the Chair of Information Management, the lab is led by Dr. Robert Jakob and Dr. Kevin O'Sullivan. It distinguishes itself from standard academic environments by prioritizing applied agentic AI—autonomous systems that move beyond chat interfaces to function within complex industrial and enterprise environments.
The lab's technical agenda is structured around several core pillars. A primary focus is what they describe as Next-Gen RAG. This research has produced frameworks like AutoRAG, which automates the configuration of retrieval pipelines, and MetaRAG, which orchestrates multiple retrieval strategies to ensure grounded outputs. By focusing on the orchestration layer, the lab addresses the inherent unpredictability of information retrieval in live production settings.
Beyond text, the lab develops Time-Series Language Models (TSLMs) and Audio-Language Models (ALMs). These projects are designed to reason over and predict temporal data in critical sectors including healthcare, energy, manufacturing, and telecommunications. This research supports their work on World Models—compressed representations of environments that enable AI systems to simulate, plan, and reason about complex dynamics before taking action.
Evaluation is a major hurdle for agent adoption, and the lab addresses this through dedicated AI Evaluation Frameworks. They develop the benchmarks and methodologies required to measure the performance, safety, and reliability of autonomous agents. This work is validated through partnerships with organizations such as Zurich Insurance, where the lab researches the deployment of agentic AI within the insurance sector. Collaborations with Stanford University and Google DeepMind ensure their research remains aligned with global technological standards.
The lab's influence extends deeply into the Swiss venture ecosystem. Through tracks like the AI Founder and AI Transformation programs, students and researchers are encouraged to turn academic projects into independent companies. This initiative has led to joint projects with various startups, including nunu.ai, which builds generally capable agents for gaming environments, and browser-use.com, a project focused on browser-based agent automation.
The lab also acts as a community hub, co-hosting events and hackathons with industry partners like Vercel and Founderful. These events provide the local ecosystem with access to high-compute resources, such as Gemini and AWS credits, and robotic platforms for testing physical agents. The team is composed of postdocs and researchers from institutions like Stanford and the National University of Singapore, creating a dense talent pool that bridges the gap between software engineering and fundamental deep learning research.
An automated configuration tool for retrieval-augmented generation pipelines.
An orchestration framework for combining multiple RAG strategies.
Agentic Systems Lab is hiring.