A build-along workshop across Finance, Healthcare, Education and beyond.
Companion repository for the Packt workshop, based on the Packt title 30 Agents Every AI Engineer Must Build by Imran Ahmad, PhD (Packt, 2026).
Format: Build-along — short concept framing, live code, then attendees build. Approx. two agents per hour. Duration: 6 hours including 1 hour of breaks (5 hours content). Trimmable to 5 hours (see Logistics). Audience: AI/ML and software engineers; technical leaders. Mixed experience welcome. Prerequisites: Laptop, Python 3.10+, Git. No API key required — every agent runs in a no-key Simulation Mode; optional keys unlock Live Mode with OpenAI, Claude, or Gemini.
The agents are sequenced to tell a story — from how a single agent reasons, to agents working together, to specialists for regulated domains, to perception of the physical world.
| # | Agent | Book | Folder | What you'll see |
|---|---|---|---|---|
| 1 | Tool-Using Agent | Ch. 7 (§7.1–7.3) | 01-tool-using-agent/ |
An LLM reasons about a request, selects the right tool from a registry, and executes it through a function-calling loop — the pattern every later agent is built on. |
| 2 | Knowledge Retrieval / RAG Agent | Ch. 6 (§6.1) | 02-knowledge-retrieval-rag-agent/ |
An end-to-end RAG pipeline — chunk, embed, FAISS index, and answers grounded only in your corpus, with source passages cited. |
| 3 | Chain-of-Agents Orchestrator | Ch. 7 (§7.4–7.6, §7.7b) | 03-chain-of-agents-orchestrator/ |
A multi-agent insurance-claims workflow (OCR → validation → risk → payout) governed by a state machine with human-in-the-loop escalation and a full audit trail. |
| 4 | Data Analysis Agent | Ch. 8 | 04-data-analysis-agent/ |
A plain-English question becomes code, runs over a dataset, and returns a live chart plus a statistical readout in seconds. |
| 5 | Verification / Fact-Checking Agent | Ch. 8 | 05-verification-fact-checking-agent/ |
A news statement is broken into claims and returned with a supported / refuted verdict and its supporting evidence. |
| 6 | Financial Advisory Agent | Ch. 14 | 06-financial-advisory-agent/ |
"RetailAdvisor": a risk profile in, a tailored portfolio recommendation out. |
| 7 | Healthcare Intelligence Agent | Ch. 13 (§13.1) | 07-healthcare-intelligence-agent/ |
Patient inputs produce a clinical decision-support output with its reasoning surfaced. |
| 8 | Education Intelligence Agent | Ch. 15 (§15.1) | 08-education-intelligence-agent/ |
An adaptive programming tutor diagnoses a learner's gap and adjusts its next explanation. |
| 9 | Vision-Language Agent | Ch. 11 (§11.1) | 09-vision-language-agent/ |
An image in, a natural-language answer out — the workshop's big visual moment. |
| 10 | Embodied Intelligence Agent | Ch. 16 (§16.1) | 10-embodied-intelligence-agent/ |
An autonomous drone mission planned under real Ottawa winter conditions, gated by a unified constraint envelope. |
Verticals covered: Finance · Healthcare · Education · Insurance · Data & Analytics · Media & Fact-Checking · Multimodal · Robotics.
- Act I — How an agent thinks (~55 min): Agents 1–2
- Break (15 min)
- Act II — Agents working together (~85 min): Agents 3–5
- Lunch (30 min)
- Act III — Domain specialists (~85 min): Agents 6–8
- Break (15 min)
- Act IV — Perception & the physical world (~55 min): Agents 9–10
- Wrap-up & Q&A (~15 min): composing these patterns into larger systems; where to go next
Each agent is one self-contained folder, so the running order and length flex easily on the day.
# Pick an agent — folders are independent, start anywhere
cd 01-tool-using-agent
# Create an environment and install that agent's dependencies
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Launch the notebook — no API key needed (Simulation Mode)
jupyter notebook 01_tool_using_agent.ipynbRun the pre-flight check at the repo root — it verifies Python, pip, venv, network, and key detection in under a minute:
python3 check_setup.py # add --full for a throwaway install testFull setup steps, FAQ, and zero-install fallbacks (Colab / Codespaces) are in PRE_CLASS_SETUP.md. Instructor email copy: PRE_CLASS_EMAIL.md.
Every notebook defaults to Simulation Mode: all LLM calls are served by a chapter-derived
MockLLM, so the full workshop runs with zero API keys and zero setup risk. Outputs are
pedagogically equivalent to Live Mode.
To run against a real model, copy the template at the repo root and set one key:
cp .env.template .env
# Edit .env — set LLM_PROVIDER (openai | anthropic | google | auto) and the matching key:
# OPENAI_API_KEY / ANTHROPIC_API_KEY / GOOGLE_API_KEYMulti-provider routing is handled by supporting/llm_provider.py, shared by all ten agents.
10-agents-workshop/
├── 01-tool-using-agent/ # Ch. 7 — tool registry + function-calling loop
├── 02-knowledge-retrieval-rag-agent/ # Ch. 6 — RAG with FAISS + cited sources
├── 03-chain-of-agents-orchestrator/ # Ch. 7 — multi-agent claims workflow + HITL
├── 04-data-analysis-agent/ # Ch. 8 — NL question → code → chart
├── 05-verification-fact-checking-agent/ # Ch. 8 — claim extraction + evidence verdicts
├── 06-financial-advisory-agent/ # Ch. 14 — RetailAdvisor portfolio recommendations
├── 07-healthcare-intelligence-agent/ # Ch. 13 — Bayesian diagnostic decision support
├── 08-education-intelligence-agent/ # Ch. 15 — adaptive tutor (BKT, IRT, SM-2)
├── 09-vision-language-agent/ # Ch. 11 — vision question answering
├── 10-embodied-intelligence-agent/ # Ch. 16 — drone mission planner (Ottawa winter)
├── supporting/llm_provider.py # shared multi-provider LLM detection
├── check_setup.py # attendee pre-flight check (--full for install test)
├── PRE_CLASS_SETUP.md # attendee setup guide + FAQ + Colab/Codespaces fallbacks
├── PRE_CLASS_EMAIL.md # instructor email copy (T-5 days / T-1 day / day-of)
├── .devcontainer/ # one-click GitHub Codespaces environment (Python 3.12)
├── .env.template # optional API keys for Live Mode
└── LICENSE # MIT
Each agent folder contains its own pre-executed notebook (outputs included), the helper
modules and data it needs, a trimmed requirements.txt, and a short README.
- A working repository with all 10 agents, runnable with OpenAI, Claude, Gemini, or a no-key simulation mode.
- A mental model for composing these patterns into larger, multi-agent systems.
- A tie-in to the book 30 Agents Every AI Engineer Must Build — each agent folder cites its chapter and sections for deeper reading.
- 5-hour cut: reduce Act IV to one agent (keep the Vision-Language Agent) or make the drone agent a recorded bonus.
- 6-hour version: full 10 agents as outlined above.
- Setup risk is low: Simulation Mode means attendees without API keys can still follow every build.
- Remote cohorts: send
PRE_CLASS_EMAIL.mdcopy 5 days out; attendees self-certify withcheck_setup.pyand followPRE_CLASS_SETUP.md(Colab and Codespaces fallbacks for locked-down laptops). Per-agent venvs are required — the agents' pinned dependencies conflict by design. - Heavy optional dependencies (e.g. torch/transformers for live vision and NLI demos) are commented out in each
requirements.txt; notebooks skip those demos gracefully without them.
MIT — see LICENSE. Code is reorganized from the book's companion repository by the author.