chabelbossa/indabax-reliable-ai-agents

Hands-on IndabaX Benin 2026 workshop on reliable AI agents, tool calling, tracing, and evals. # Building Reliable AI Agents **Tool Calling, Orchestration and Evaluation in Practice** Deep Learning IndabaX Benin 2026 - Educational lab by BOSSA Chabel In this 50-minute hands-on lab, you will build a small agent that chooses and calls deterministic tools, take it from a fragile `5 / 8` baseline to `8 / 8`, and add validation, controlled errors, tracing, and behavior-based evaluations. > Making an agent work is easy. Knowing when it fails is harder. ## What you will learn - How a tool-using agent differs from a chatbot. - What an LLM function call contains and who executes it. - How a bounded agent loop turns tool results into a final answer. - Why generated arguments must be validated. - How traces and a small eval suite make failures inspectable. ## Prerequisites - Basic Python: functions, dictionaries, and JSON. - Python 3.10+ for local execution, or Google Colab. - A Gemini API key only if you want live mode. Mock mode is fully offline. No Docker, database, GPU, vector store, or agent framework is required. ## Fastest start: Google Colab 1. Open the participant notebook. 2. Run the setup cell. It defaults to `MODE: MOCK` and needs no key. 3. Complete the three short TODOs. 4. If you are blocked, open the solution notebook. ## Local quick start ```bash git clone github.com cd indabax-reliable-ai-agents python3 -m venv .venv source .venv/bin/activate python -m pip install -r requirements.txt python -m evals.run_evals ``` Expected result: `8 / 8 passed`. ## Live Gemini mode The default is deliberately offline and deterministic: ```bash export LLM_MODE=mock ``` For a real function-calling run, create a key in Google AI Studio, then inject it into your shell without putting it in a file or notebook output: ```bash export LLM_MODE=gemini export GEMINI_API_KEY='your-key-here' ``` The selected mode is always printed. The workshop never silently presents mock output as a live model response. Fr …