AI Agents Bootcamp workbench

Companion repo for the Udemy course AI Agents Bootcamp (listing 6521157).

Which branch do I need? The course was rebuilt in September 2026, and the branch you are reading is the rebuilt course: every lecture names a file under labs/, and each lab comes as a plain .py script and an .ipynb twin, so you can work in an editor or in a notebook. It is written as a real repo on purpose, because that is how agent products are built at work: modules, tests, config, evals and a deploy door, not one long notebook. Took the 2025 edition? Your notebooks are on the branch original-2025, exactly as you had them: git fetch && git checkout original-2025. That branch is archived and no longer updated.

Three builds:

• DataFlow desk: a LangGraph support desk with retrieve, a parked refund, citations, and a refuse line.
• Research and report agent: a plan, files, isolated subagents, and a sourced report on disk.
• TalentFlow pipeline: map-reduce over resumes, a job description, and email templates.

Four files a clone needs: README (this file), smoke (deploy/smoke.py), golden set (eval/golden.jsonl), the FastAPI door (dataflow/serve/app.py).

Clone and test (no key)

git clone https://github.com/pragatidev/AIAgentsBootcamp.git
cd AIAgentsBootcamp
py -3.13 -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
pytest -q

Use Python 3.13 from python.org's standalone installer; 3.11 and 3.12 also work, but python.org no longer gives them regular bug fixes or binary installers. The Python docs (read 2026-10-04) recommend the Python install manager and mark the classic installer and the classic py launcher as deprecated since 3.14. If PowerShell says Activate.ps1 cannot be loaded because running scripts is disabled on this system, run Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser once (the fix the Python venv docs give) and activate again, or use cmd. On our test machine with a warm pip cache the install took 247.2 seconds and 1.82 GB on Python 3.13 (278.8 seconds and 1.9 GB on 3.11), and a first install downloads everything, so expect it to take longer; PyTorch comes in through sentence-transformers. macOS / Linux: python3 -m venv .venv and source .venv/bin/activate. pytest -q must exit 0 with no cloud key and print the N passed line. Copy .env.sample to .env (copy on Windows, cp on macOS / Linux) when you add a key or change a model. Live model calls skip without a key or Ollama. The browser labs (Playwright) need a browser that pip does not install: run playwright install chromium once; until then their eight tests skip and say so. Full steps: INSTALL.md. Errors and fixes: TROUBLESHOOTING.md.

What the desk refuses

The DataFlow desk refuses when the knowledge base has nothing. It does not invent a policy. It does not write a refund until a named reviewer resumes the parked card.

For labs that call a model, install Ollama and run ollama pull qwen3:8b. Models live in config.py. Never commit .env. Capstones: labs/19_dataflow/starter, labs/19_research/starter, labs/19_talentflow/starter. Part 16 compares agent frameworks: run pip install -r requirements-frameworks.txt once (CrewAI and AutoGen get their own venv, the steps are in that file).

Layout

techcorp/           IT desk (LangChain)
dataflow/           customer desk, graphs, RAG, MCP, FastAPI
talentflow/         resumes, job, email templates
research_agent/     research and report package
eval/golden.jsonl   golden set
deploy/smoke.py     clone smoke
labs/               teaching scripts plus .ipynb twins
tests/              green without a key
config.py           the only place a model id lives
src/                the model door (llm.py), the repo finder (paths.py), helpers the Section 2 labs use
harness/            the DataFlow harness: guides, sensors, permissions, and recorded runs
scripts/            repo tools, such as make_twins.py, which rebuilds the notebook twins
evals/              one line that points to eval/harness.py
exercises/          a starter for one coding exercise (an unknown order id)
solutions/          the solution to that exercise
prompts/            the DataFlow system prompt
business/           a DataFlow offer and price, and TalentFlow versus a recruiting ATS
career/             PM pages (spec, metrics, cost model) and a QA test plan for DataFlow
docs/               CURRENCY.md: check that a model id is live before you use it
ops/                tracing/runs/, where the DataFlow service writes its request log
.github/            a workflow that runs the eval gate on every pull request
conftest.py         keeps pytest out of the lab scripts

RAG (Part 8)

Local default embedder is nomic-embed-text on Ollama. Chat is qwen3:8b.

python -m dataflow.rag.faiss_index                 # build the FAISS index (dataflow/data/faiss_index, gitignored)
python labs/08_04_06_portfolio_knowledge_desk.py   # three tickets: a cited policy, an unknown that refuses, an order id that does not retrieve
python labs/08_04_05_rag_evals.py                  # naive vs agentic evals on eval/questions.jsonl

First lab: labs/00_03_setup_check.py. Twins: python scripts/make_twins.py. License MIT.