IE 685: Large Language Models and Agents (University of Mannheim, FSS 2026)
by University of Mannheim, Data and Web Science Group
A master's course on LLM agents, RAG and context engineering, with public slides and LangChain exercises.
Overview
IE 685: Large Language Models and Agents is a 3 ECTS master's lecture from the Data and Web Science Group at the University of Mannheim, taught in spring 2026 (FSS 2026) by Ralph Peeters, Christian Bizer and Aaron Steiner. It pairs with IE 686, a 3 ECTS project course. The course page links seven lecture decks: introduction to LLMs and agents, post-training and efficient adaptation, prompt engineering, LLM agents and tool use, retrieval augmented generation, context engineering, and security and safety. Five exercises accompany them, covering Hugging Face Transformers, LLM workflows with LangChain, agent programming with LangChain, RAG workflows with Tavily and LangChain, and a deep agent for context engineering. The RAG deck, dated March 10, 2026, walks through chunking, embedding and indexing, retrieval and re-ranking, query expansion and decomposition, generation, workflow patterns, deep research agents, and RAG metrics and benchmarks. The context engineering deck, dated March 16, 2026, covers context rot, token costs and prompt caching, then four techniques: write, select, compress and isolate. The security deck, dated April 21, 2026, covers prompt injection, jailbreaking, data exfiltration and malicious tool use, with guardrails, sandboxing and access control as mitigations. Enrollment is limited to 50 Mannheim students, and outside learners get the public materials only. No lecture videos are listed for this edition, so plan to learn from slides and code.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- ~18 classroom hours in the original format (7 lectures and 5 exercises at 90 minutes each); self-paced from slides and code, no videos listed
- Provider
- University of Mannheim, Data and Web Science Group
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Use Hugging Face Transformers in exercises alongside lectures on post-training and efficient model adaptation
- ✓Build LLM workflows and program tool-using agents with LangChain in the course's hands-on exercises
- ✓Build a RAG pipeline covering chunking, embedding and indexing, retrieval, re-ranking, query expansion and generation
- ✓Design RAG workflows and deep research agents, then assess them with evaluation metrics and benchmarks
- ✓Manage agent context with write, select, compress and isolate techniques, using a deep agent exercise
- ✓Account for context rot, token costs and prompt caching when designing what an agent's context holds
- ✓Recognize prompt injection, jailbreaking, data exfiltration and malicious tool use, and apply guardrails, sandboxing and access control
Highlights
- •Most lecture weeks pair a slide deck with a downloadable exercise, so each concept comes with code to run
- •Context engineering gets a full lecture and its own deep agent exercise instead of a passing mention inside prompting
- •The materials are recent, with decks versioned between March and April 2026, and the page also links published student evaluation results for the spring 2026 edition
- •The course changes between years: its 2025 introduction deck names LangGraph and AutoGen as exercise tools, while the 2026 exercises use LangChain and Tavily
- •Limits to plan around: no lecture videos are listed, and evaluation appears in the topic list without a standalone deck, though the RAG lecture covers RAG metrics and benchmarks
Who It’s For
Best For
- ✓Developers with Python and basic ML who want a structured, university-level path into LLM agents and RAG
- ✓Self-learners who are comfortable learning from slides and code without recorded lectures
- ✓Instructors and team leads looking for a compact agents curriculum with exercises to adapt
Prerequisites
- •Python programming skills, as stated on the course page
- •Basic machine learning concepts and techniques, as stated on the course page
FAQ
What is IE 685: Large Language Models and Agents (University of Mannheim, FSS 2026)?
IE 685 is a University of Mannheim master's course for developers with Python and basic machine learning who want a structured route into LLM applications. Working through its slides and exercises prepares you to build LangChain agents, RAG workflows and context-managed deep agents, and to think through their security.
Is IE 685: Large Language Models and Agents (University of Mannheim, FSS 2026) free?
IE 685: Large Language Models and Agents (University of Mannheim, FSS 2026) is free to access.
What level is IE 685: Large Language Models and Agents (University of Mannheim, FSS 2026) for?
IE 685: Large Language Models and Agents (University of Mannheim, FSS 2026) is aimed at a intermediate audience. Recommended background: Python programming skills, as stated on the course page, Basic machine learning concepts and techniques, as stated on the course page.
How long does IE 685: Large Language Models and Agents (University of Mannheim, FSS 2026) take?
Expect roughly ~18 classroom hours in the original format (7 lectures and 5 exercises at 90 minutes each); self-paced from slides and code, no videos listed. Most learners work through it at their own pace.
What will I learn from IE 685: Large Language Models and Agents (University of Mannheim, FSS 2026)?
You'll learn: Use Hugging Face Transformers in exercises alongside lectures on post-training and efficient model adaptation; Build LLM workflows and program tool-using agents with LangChain in the course's hands-on exercises; Build a RAG pipeline covering chunking, embedding and indexing, retrieval, re-ranking, query expansion and generation; Design RAG workflows and deep research agents, then assess them with evaluation metrics and benchmarks; Manage agent context with write, select, compress and isolate techniques, using a deep agent exercise; Account for context rot, token costs and prompt caching when designing what an agent's context holds; Recognize prompt injection, jailbreaking, data exfiltration and malicious tool use, and apply guardrails, sandboxing and access control.
Topics
Sources
This page was written from 5 sources, 1 on domains other than uni-mannheim.de.