🤖 Master AI Agentic Engineering

Build & Deploy Autonomous AI Agents — a 6-Week Journey

Five leading agent frameworks. Six real projects. One repository.

---

🧭 Overview

This repository is my complete record of a 6-week, hands-on journey building autonomous AI agents. Each week introduces a framework through a series of labs (learn the concepts) and culminates in a real, runnable project (apply them).

Week	Framework	Capstone Project	Highlight
1	Raw LLM APIs + Tools	🧑‍💼 Career Digital Twin	An AI that is you
2	OpenAI Agents SDK	🔬 Deep Research Assistant	4-agent research pipeline
3	CrewAI	👷 Engineering Team (+4 crews)	A team that ships software
4	LangGraph	🦾 Sidekick	Self-evaluating co-worker
5	AutoGen	🏭 Agent Factory	Agents that create agents
6	Model Context Protocol	📈 Autonomous Trading Floor	4 traders, live & autonomous

---

🧰 Tech Stack — Everything Used in the Course

🧠 Agent Frameworks

🤖 Models & Providers

🛠️ Tools, Data & Integrations

⚙️ Language & Environment

📋 Full capability matrix (click to expand)

Category	What's covered
Agent patterns	tool calling · agents-as-tools · handoffs / delegation · planner→worker→evaluator loops · hierarchical (manager) crews · agents that generate agents
Tools & integrations	web search (Serper / Brave / hosted WebSearchTool) · Playwright browser automation · Python REPL · Docker code execution · file management · Wikipedia
State & memory	Pydantic structured outputs · LangGraph checkpointers · CrewAI short / long-term / entity memory · SQLite & ChromaDB persistence
Multi-model	OpenAI · Anthropic · Ollama · DeepSeek · Gemini · Grok (via OpenRouter)
Ops & delivery	observability / tracing · Gradio UIs · Pushover & SendGrid notifications · gRPC distributed runtime · MCP servers / clients / resources

---

🚀 Quick Start

This project uses uv for dependency and environment management.

# 1 — Install uv (if needed)
pip install uv

# 2 — Sync the environment from pyproject.toml / uv.lock
uv sync

# 3 — Add your secrets (see below)
#     create a .env file in the project root

🔑 Environment variables (click to expand)

Create a .env file in the project root (agents/) with the keys you need:

OPENAI_API_KEY=...
ANTHROPIC_API_KEY=...
SERPER_API_KEY=...          # web search (CrewAI)
PUSHOVER_TOKEN=...          # push notifications
PUSHOVER_USER=...
SENDGRID_API_KEY=...        # email (Deep Research)

# Week 6 — optional multi-model + market data
DEEPSEEK_API_KEY=...   GOOGLE_API_KEY=...   GROK_API_KEY=...
OPENROUTER_API_KEY=... BRAVE_API_KEY=...    POLYGON_API_KEY=...

📓 Notebooks (*.ipynb) run lab-by-lab · 🖥️ Apps launch with uv run <file>.py · 👥 CrewAI projects run with crewai run from inside the project folder.

---

📅 The 6-Week Journey

🟢 Week 1 — Foundations

📁 1_foundations  ·  🧱 OpenAI / Anthropic / Ollama SDKs · Gradio · Pushover

📖 What I learned Calling LLMs directly via raw APIs, comparing multiple models (OpenAI, Claude, local via Ollama). Designing prompts, system messages, and multi-step LLM workflows (one model feeding the next). Tool calling from first principles — hand-written JSON tool schemas + an execution loop. Reading a PDF as context and shipping a chat UI with Gradio.

🚀 Project · Career Digital Twin A chatbot (app.py) that acts as me — answering questions about my career from my LinkedIn PDF + a personal summary. Two tools push to my phone via Pushover: one records an interested visitor's email, the other logs any question it couldn't answer. The foundational pattern — LLM + context + tools + UI — that every later framework builds on.

---

🔵 Week 2 — OpenAI Agents SDK

📁 2_openai  ·  🧱 OpenAI Agents SDK · SendGrid · Tracing

📖 What I learned Core primitives: Agent, Runner, function_tool, trace (observability dashboard). Agents as tools — wrapping one agent so another can call it. Handoffs — delegating control to a specialized agent. Structured Outputs with Pydantic for typed, validated results. A Sales Development Rep workflow: a planning manager orchestrating writer agents + an email handoff.

🚀 Project · Deep Research Assistant A multi-agent pipeline (deep_research/) behind a Gradio UI: 🧠 planner → N structured search terms 🔎 search → parallel web search + summaries ✍️ writer → long structured markdown report 📧 email → clean HTML, sent via SendGrid A ResearchManager orchestrates the flow and streams live status to the UI.

---

🟣 Week 3 — CrewAI

📁 3_crew  ·  🧱 CrewAI (Agent/Task/Crew) · YAML config · Serper · Pydantic · Memory · Docker

📖 What I learned

• The CrewAI model: agents & tasks declared in agents.yaml / tasks.yaml, assembled by a @CrewBase class.
• Sequential vs hierarchical processes (a manager agent that delegates).
• Tools (SerperDevTool, custom), Docker code execution, context passing between tasks, Pydantic outputs, and persistent memory (short-term RAG · long-term SQLite · entity).
• Mixing models per-agent (GPT-4o lead, Claude Sonnet coder/judge).

🚀 Projects · Five Crews

Crew	What it does
💻 coder	A Python-developer agent that writes & runs code in a safe Docker sandbox to solve an assignment.
⚖️ debate	A debater argues both sides of a motion; a judge (Claude) decides the winner.
📊 financial_researcher	A researcher (web search) + analyst produce a polished company report.
📈 stock_picker	Hierarchical crew with a manager: finds trending companies, researches them, picks the best stock — with memory + push notifications.
⭐ engineering_team	The capstone — a full software team: a Lead designs → a Backend Engineer codes (Docker) → a Frontend Engineer builds a Gradio UI → a Test Engineer writes unit tests. Requirements → working app.

---

🟠 Week 4 — LangGraph

📁 4_langgraph  ·  🧱 LangGraph · LangChain tools · Playwright · Gradio

📖 What I learned The graph model: a typed State (with add_messages reducers), nodes, edges, and conditional edges for routing. Compiling with a checkpointer (MemorySaver) for persistent, thread-based memory. A rich LangChain toolset: Playwright browser, file management, Python REPL, Wikipedia, Serper search. The worker → tools → evaluator loop.

🚀 Project · Sidekick A self-evaluating co-worker (sidekick.py + app.py). You give it a request and a success criterion. A worker agent uses tools to attempt the task; an evaluator agent judges the result against your criterion — approving it, or routing back with feedback until the bar is met (or it needs to ask you a question).

---

🔴 Week 5 — AutoGen

📁 5_autogen  ·  🧱 AutoGen AgentChat + Core · gRPC distributed runtime

📖 What I learned AutoGen AgentChat — AssistantAgent, conversations, tool use. AutoGen Core — the actor model: RoutedAgent, @message_handler, message passing, runtime registration. Running agents on a distributed gRPC runtime so many run concurrently.

🚀 Project · Agent Factory A self-expanding world of entrepreneurs (world.py). A Creator (creator.py) takes an agent template (agent.py), asks an LLM to generate a brand-new agent with a unique personality, writes it to a file & registers it live. Each agent invents a business idea and may bounce it off another agent to refine it — 20 agents spawning ideas on the gRPC runtime.

---

⚫ Week 6 — MCP (Model Context Protocol)

📁 6_mcp  ·  🧱 MCP (clients + servers) · OpenAI Agents SDK · multi-model · SQLite · Gradio

📖 What I learned

• What MCP is: a standard protocol connecting agents to tools, data & resources via MCP servers.
• Consuming MCP servers in the OpenAI Agents SDK (MCPServerStdio) — incl. community servers (fetch, Brave, memory, Playwright).
• Writing my own MCP servers (accounts_server.py, market_server.py, push_server.py) and a custom client.
• Exposing not just tools but resources (reading an account / strategy over MCP).
• Running agents across many models (GPT · DeepSeek · Gemini · Grok via OpenRouter).

🚀 Project · Autonomous Trading Floor

A continuously-running simulated hedge fund (trading_floor.py):

• 🧑‍💼 Four traders with distinct strategies — Warren (patience) · George (bold) · Ray (systematic) · Cathie (crypto) — each optionally on a different LLM.
• 🔬 Each has a Researcher sub-agent (web fetch + Brave search + persistent memory over MCP), trading against a market via MCP servers and managing a real Account (balance, holdings, transactions, P&L in SQLite).
• ⏰ A scheduler wakes the floor every N minutes during market hours — fully observable via tracing + a Gradio dashboard (app.py).

🏆 The grand finale — multi-agent orchestration + tools + memory + the MCP standard + multi-model, running autonomously.

---

⭐ 6 frameworks · 6 projects · 1 journey from raw API calls to fully autonomous agents.

Built with the OpenAI Agents SDK · CrewAI · LangGraph · AutoGen · MCP