Agentic Artificial Intelligence and Multi-Agent Systems

Hands-on course · Fundacion AI Granada Research & Innovation 13, 14 and 16 July 2026 · 10:00-12:00 · Online (live sessions)

Taught by Montevive AI · Chema Robles (chema@montevive.ai)

Course notebooks

Notebook	Session	Open in Colab
00-environment-check.ipynb	Before you start
session-1-agent-fundamentals.ipynb	Mon 13 · Agentic AI Fundamentals
session-2-multi-agent-systems.ipynb	Tue 14 · Multi-Agent Systems
session-3-production-mcp.ipynb	Thu 16 · Agents in Production and Frontiers

Before the first session

1. Set up the environment (local or Colab, see below).
2. Get an API key: the course key for the Montevive LLM proxy (given at the start of the training), or your own provider key (Anthropic, OpenAI or Google).
3. Run 00-environment-check.ipynb end to end and check that it shows 🟢.

If anything fails: chema@montevive.ai.

Option A · Local environment (recommended)

Requirements: Python 3.11+ (3.12 recommended), git.

git clone https://github.com/montevive/agentic-ai-course.git
cd agentic-ai-course

python3 -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

pip install -r requirements.txt

cp .env.example .env               # and fill in your API keys

jupyter lab                        # or open the folder in VS Code

With conda:

conda create -n agentic-course python=3.12 -y
conda activate agentic-course
pip install -r requirements.txt

Option B · Google Colab

Use the badges in the table above. In Colab:

• The first cell of each notebook installs the dependencies automatically.
• Add your key in the Secrets panel (🔑 icon in the sidebar): LITELLM_API_KEY with the course key (recommended), or your own ANTHROPIC_API_KEY / OPENAI_API_KEY / GOOGLE_API_KEY. Enable "Notebook access".

LLM access: the Montevive proxy (recommended)

The course runs on the Montevive LLM proxy (llm.montevive.ai). You get one key and set it as LITELLM_API_KEY; every notebook then routes all three providers through the proxy automatically — no other change needed, model strings like anthropic:claude-sonnet-4-6 keep working.

Why a proxy: it only allows the course models, so you can't accidentally run an expensive model, and it shares one daily budget. This is also a real example of a governance guardrail — model allowlisting, per-key budget and rate limits — the same "secure & governed AI" lens the course applies to compliance in Session 3.

Cheap (default in exercises)	Mid tier
claude-haiku-4-5	claude-sonnet-4-6
gpt-5-mini	gpt-5
gemini-flash-latest	gemini-pro-latest

Any other model returns 403 (intentional). The key is shared by the class: ~20 USD/day budget, 300 req/min, and it expires 2026-07-18. Prefer the cheap tier and avoid unbounded agent loops. A 429 means wait a few seconds and retry.

Without the proxy: leave LITELLM_API_KEY empty and set your own provider key instead (ANTHROPIC_API_KEY recommended, or OPENAI_API_KEY / GOOGLE_API_KEY) — the notebooks then call the providers directly. One key is enough; estimated spend across the three sessions is under 2-3 EUR with the default cheap models.

Repository structure

├── 00-environment-check.ipynb          Environment check (run before S1)
├── 101-real-world-agent-systems.md     Field guide: 101 real agent systems across 13 domains, fact-checked
├── 101-open-source-agent-systems.md    Companion: 101 open-source agents you can clone and adapt, API-verified
├── session-1-agent-fundamentals.ipynb  From LLM to agent · anatomy · Pydantic AI, smolagents, OpenAI Agents SDK
├── session-2-multi-agent-systems.ipynb Coordination · CrewAI vs LangGraph · context, memory and RAG
├── session-3-production-mcp.ipynb       MCP · guardrails and traceability · end-to-end pipeline · frontiers
├── data/                               Mini-corpus of abstracts (fictional) for RAG and MCP
├── assets/                             Figures, logos and banners used in the notebooks (SVG)
├── mcp_servers/
│   └── corpus_server.py                FastMCP server used in Session 3
├── a2a_servers/
│   └── corpus_agent.py                 A2A server (Agent Card + JSON-RPC) used in Session 2
├── n8n/
│   ├── agent-prototype.json            Importable n8n workflow mirroring the Session 1 agent
│   ├── multi-agent-research.json       Supervisor + sub-agents as tools (Session 2)
│   └── README.md                       How to import and run them (n8n Cloud or Docker)
├── skills/
│   └── literature-review/              Example Agent Skill (SKILL.md) used in Session 3
├── requirements.txt                    Pinned and verified dependencies
└── .env.example                        Environment variables template

Program

Session	Date	Content
S1 · Agentic AI Fundamentals	Mon 13 Jul	From LLM to agent (ReAct, tool use) · visual prototyping with n8n (importable workflow in n8n/) · anatomy of an agent · live implementation with Pydantic AI and smolagents · contrast with OpenAI Agents SDK
S2 · Multi-Agent Systems	Tue 14 Jul	Coordination patterns and A2A (live demo) · supervisor pattern in n8n (agents as tools) · CrewAI vs LangGraph vs Microsoft Agent Framework · multi-agent workshop · context engineering, memory (Mem0) and RAG (ChromaDB, cosine, hybrid BM25)
S3 · Production and Frontiers	Thu 16 Jul	Model Context Protocol (FastMCP) · Agent Skills (SKILL.md) · guardrails, prompt injection, PII filtering (Presidio) and traceability (Langfuse, NVIDIA NeMo Agent Toolkit) · from guardrails to compliance (GDPR, AI Act, NIS2, ISO 27001) · end-to-end multi-agent pipeline · trends 2026-2027

Looking for real-world examples? 101 real-world agent systems is a fact-checked field guide of production and research agents across 13 domains, mapped to the coordination patterns taught in Session 2. Its companion, 101 open-source agent systems, lists another 101 agents whose code you can clone and adapt, every repository verified live with stack and star snapshot.

About Montevive AI

Montevive AI is an AI engineering company based in Granada. Our mission is to help businesses apply AI securely and aligned with compliance — GDPR, NIS2, ISO 27001, AI Act — with data protection present from the design. That lens runs through the whole course: look for the "Secure & governed AI" asides in each session, the live prompt-injection demo (Prompt Injection Indirecto: Robando API Keys de Agentes IA, Spanish), the PII/PFI/PHI filtering exercise and the guardrails-to-regulation map in Session 3.

License and use

Training material produced by Montevive AI for Fundacion AI Granada. The abstracts in data/ are synthetic (fictional papers and authors). You may reuse the exercise code in your research and teaching by citing the source.