Practical Course Deep Reinforcement Learning & Agentic AI Systems

(IN0012, IN2106)

Supervision: Prof. Dr. Martin Bichler, Kassian Köck.

Agentic AI systems study the interaction of multiple autonomous decision-making entities that operate in shared environments and adapt their behaviour over time. This area lies at the intersection of artificial intelligence, machine learning, multi-agent systems, and economic decision theory, and has gained renewed attention in recent years due to advances in deep reinforcement learning, large-scale computation, and large language models. At the same time, autonomous agents and bots are increasingly deployed in real-world economic and digital settings such as online markets, negotiation platforms, and automated decision systems, raising new challenges related to learning dynamics, coordination, stability, robustness, and strategic behaviour.

Previous Knowledge Expected

Students are expected to have strong programming skills in Python. In addition, the following background is required:

• a solid foundation in algorithms, data structures, and software engineering,
• basic knowledge of artificial intelligence or machine learning,
• an interest in multi-agent systems, distributed systems, or market-based models,
• the willingness to independently familiarise themselves with new frameworks, libraries, and APIs, and
• prior successful completion of at least one of the following courses: Business Analytics and Machine Learning (IN2028), Machine Learning (IN2064), or Introduction to Deep Learning (IN2346).

Participation in the pre-course meeting is strongly recommended.

Module Classification

• IN0012 (Practical course in the Bachelor program Informatics).
• IN2106 (Practical course in the Master program Informatics).
• For all other programs: Please check first whether this practical course fits in your curriculum.

Teaching and Learning Method

In this practical course, students design and implement agent-based market systems in which autonomous software agents interact in economic environments such as auctions, negotiations, matching markets, and platform-based marketplaces. These market settings serve as an application domain to study and evaluate agentic AI systems that combine deep reinforcement learning with modern large-language-model-based agents.

Students implement agents that learn and adapt their strategies using reinforcement learning methods (e.g. policy-based or value-based approaches) and integrate these with symbolic and language-based decision components. Learning-based components may be realised using modern numerical and machine learning frameworks such as JAX, while high-level agent behaviour, communication, coordination, and tool use are implemented using contemporary agent frameworks (e.g. Microsoft Agent Framework, AutoGen, and related agent-to-agent communication paradigms).

Participants design market mechanisms and specialized agent roles (e.g. buyers, sellers, auctioneers, intermediaries), integrate external tools and stateful components (memory, retrieval, tool use), and analyse system behaviour under different informational and strategic assumptions. A central theme of the course is the interaction between learned policies, language-based reasoning, and explicitly defined market rules.

Beyond functional correctness, the course emphasizes the experimental evaluation of learning dynamics, convergence, efficiency, and stability in multi-agent systems. Additional aspects include robustness, scalability, and security issues in agentic systems (e.g. prompt injection, unreliable tool outputs, or strategic manipulation).

The course follows a project-based format with students working in small teams. Results are presented in the form of executable prototypes, controlled experimental evaluations, and a written project report.

Course Objective

After successful completion of the course, students will be able to:

• design and implement autonomous software agents with clearly defined roles and interaction protocols,
• realize agent-based market systems and market mechanisms in practice,
• apply modern agent frameworks for tool use, memory, and multi-agent coordination,
• experimentally analyze complex multi-agent systems and define suitable metrics for efficiency, stability, and robustness,
• identify and address common failure modes and security risks in agentic systems, and
• clearly document and present technical designs, system architectures, and experimental results.

Course Criteria and Registration

The practical course is primarily aimed at BSc and MSc students in Computer Science and adjacent degree programs such as Information Systems, Data Science, or Mathematics.

All students (from all schools) must apply via the School of CIT's matching system in the "Practical Course SS26" matching instance. (Additional information here.)

To improve their chances of being selected, students may send a current transcript of records to Kassian Köck via email within the registration deadline of the matching system. Students may optionally include further information in this email, such as their motivation for participating, relevant experience outside of university courses, or relevant courses taken or planned that are not yet listed on the transcript.

Additional Ressources

Interested students are encouraged to apply early for the free student subscriptions to Google Gemini Pro and/or GitHub Copilot. While the offer is currently available, no official end date has been announced, and the application process may take some time to be approved.