Multi-Agent AI Lab

Course Details

  • Course Number: IN2106
  • Full Title: Multi-Agent AI Lab: LLM Agents, Retrieval and Resource-Aware Cooperation
  • Format: Master Practical Course
  • Instructors: Matthias Grabmair, Julia Schaumeier
  • Language: English
  • SWS: 6
  • ECTS: 10

Information Session

Watch the recorded information session for our winter semester 2026/27 lab courses:

Content Outline

This practical course addresses multi-agent systems and modern AI systems that combine heterogeneous agents, LLMs, retrieval, communication, planning, and resource-aware decision making. The central project is a configurable, tournament-style environment in which teams of agents solve knowledge-intensive tasks over a fixed corpus.

Agents must retrieve and aggregate information, exchange partial findings, decide when to use an LLM, manage token and compute budgets, and submit final answers with supporting evidence. The environment combines cooperative elements within each group with competitive elements between groups through a shared benchmark or leaderboard.

Students work in groups of three to four and develop their systems across three milestones. Each student contributes at least one agent or major system component, preferably based on a distinct agent paradigm. The environment becomes gradually harder over the semester — for example through partial observations, limited or noisy communication, distractor evidence, scarce token budgets, runtime limits, and hidden evaluation settings. Teams meet regularly with a mentor and present their progress at each milestone.

Requirements

Students should have programming experience, preferably in Python, and basic knowledge of artificial intelligence and machine learning. Prior experience with multi-agent systems, NLP, information retrieval, deep learning, reinforcement learning, APIs, or software engineering is helpful but not required; students are expected to independently acquire missing background during the first weeks.

For LLM-based components, students use available university computing resources. In addition, the research group provides controlled access to a course-specific LLM inference service (subject to availability, quotas, and fair-use rules). Small open-weight models may also be run locally on suitable personal hardware.

Learning Outcomes

After completing the course, students can:

  • Explain and compare major agent paradigms
  • Implement heterogeneous multi-agent systems
  • Design LLM-based tool-using agents
  • Build retrieval-augmented knowledge workflows
  • Reason about token and compute budgets
  • Design controlled experiments
  • Evaluate systems under environmental variation
  • Present results in a scientific report

Assessment

Grades are based on three milestones. For each, teams submit progress documentation and present their work; individual contributions are documented through code, an individual agent design note, and milestone reflections.

  • First milestone (e.g. project design, literature survey, retrieval-enabled baseline, first agent prototypes): 20% work, 10% presentation
  • Second milestone (e.g. functioning multi-agent prototype with communication, LLM integration, knowledge aggregation, resource logging): 20% work, 10% presentation
  • Final milestone (e.g. tournament-ready system, experiments, ablations, final report and demo): 30% work, 10% presentation

A leaderboard or tournament may be included, but grading primarily reflects scientific and engineering quality.