Workshop: Build a Full AI Agent From Scratch

​Build a Full AI Agent From Scratch: RAG + MCP + Chat

A hands-on AI workshop in collaboration with the Department of Informatics at UiB.

Free and open to everyone.

Agents are moving from hype to workhorse, and companies are hiring people who can build them properly. The difference between a demo and a product comes down to good retrieval, correct routing and sensible context management. Without those, an agent is slow, expensive or simply wrong.
In one evening you’ll build the core of a real agent system yourself. The skills go straight into your thesis, your course projects, your side projects and your CV.


What you’ll learn

• ​Embeddings in practice: choose a model, understand similarity metrics (cosine vs L2) and build a FAISS index that scales. If the agent can’t find the right document quickly, it fails in the real world.

• ​MCP-driven agent selection: describe areas of responsibility semantically and route each request to the right agent. Less manual wiring, fewer mis-routings, easier to extend.

• ​The full RAG loop with chat: keep multi-turn context, retrieve the right knowledge, synthesize an answer, and fall back gracefully when the agent is unsure. This is the foundation of robust agents in support, internal tools and knowledge apps.

​Agenda (hands-on)

• ​Basic vector search (20 min): embeddings → FAISS → top-k. You build ask_rag(query, k) and test it.

• ​MCPs & agent selection (40 min): define three MCPs (ticketing, user_management, notifications). select_mcp(query) picks the right agent via cosine similarity.

• ​Refreshments (30 min)

• ​Full RAG + chat (50 min): chat history, dynamic MCP selection, document retrieval inside the MCP, top-k aggregation and answer generation. Bonus: fallback agent, multi-agent collaboration, LLM summarization.

​You don’t need any prior experience with agents, RAG or MCP. If you’re comfortable with basic Python, you’re good to go. We’ll guide you through the rest, and mentors are on hand if you get stuck.

What you need

• ​A laptop with Python 3.10+

• ​An editor you like (VS Code works great)

• ​About 3 hours: a short intro, then lots of building

​Extra challenges (for the hungry)

• ​Dynamic chunking of long texts

• ​Multi-agent orchestration (retrieve → summarize → enrich)

• ​Visualize the flow: Query → MCP → FAISS → Response

• ​Benchmark MCP routing and top-k strategies

​What you take with you
A working agent you can connect to your own datasets and APIs, for example your course notes, a thesis corpus or a hobby project, plus the know-how to extend it.

Spots are limited, so sign up early.
Bring your laptop, and let’s build agents that actually deliver.