🏋️‍♂️ Agentic Coding Fitness @ Rust Tech Bar

Welcome to the repository for the Agentic Coding Fitness event series, hosted weekly at Rust Bar, Ban Tad Thong!

This repository contains all the code, tools, and examples built during our hands-on "Vibe Coding" sessions. It serves as a living codebase demonstrating how to transition from basic AI API calls to building sophisticated, multi-agent systems and real-world IoT integrations.

Event Details: Luma Event Page

• When: Every Tuesday, 18:00 – 20:00
• Where: Rust Bar, Ban Tad Thong (Bangkok)

🤖 What is Agentic Coding Fitness?

Think of this as a "fitness center" for your coding brain—but instead of lifting weights, we are building AI muscle muscle memory. We focus on Agentic AI: moving beyond simple prompt-and-response mechanisms to build AI that can think, plan, decide, and collaborate using multi-agent systems.

We emphasize a practice-first approach (Vibe Coding). No long lectures, just shipping workable solutions that interact with the real world!

📚 Catching up? Install the Bootcamp Plugin

Missed a session or want to review at your own pace? We packaged the entire course (weeks 2–18) into a shareable Claude Code plugin — 20 bite-sized skills (one per concept) that teach the idea, show runnable code (pointing at the real weekN/ files here), and walk you through a hands-on $0 lab (a tiny MockLLM, or fully offline checkpoints, so you need no API key to start). Just ask Claude in plain English and the right skill loads automatically.

Install (run these in any Claude Code session):

/plugin marketplace add kwarodom/agenticaicodingfitness
/plugin install agentic-coding-fitness@agentic-coding-fitness

Then try: "Recap the whole course and tell me which skill to start with."

🔄 Already installed? Pull the latest version mid-session

We ship new skills as the course grows (we're on v2.2.0 — 20 skills). To grab the newest version without restarting, run these three in your current session:

/plugin marketplace update agentic-coding-fitness                 # 1. refresh the catalog from GitHub
/plugin install agentic-coding-fitness@agentic-coding-fitness     # 2. fetch the latest version
/reload-plugins                                                   # 3. activate the new skills now

There's no separate /plugin update command — reinstalling pulls the latest version from the refreshed marketplace. Step 1's argument is the marketplace name (agentic-coding-fitness), not the GitHub repo.

Prefer clicking? Run /plugin for the interactive manager: Marketplaces tab → select agentic-coding-fitness → Update, then Installed tab → select the plugin → Reinstall, then /reload-plugins. (You can also toggle Enable auto-update on the marketplace so new versions are fetched at startup.)

Covers: LLM basics · tool use · agent loops (now incl. the Week 18 Claude Agent SDK production loop) · MCP & skills · RAG · multi-agent systems · production & observability · agent evaluation/CI · knowledge-graph memory · production GraphRAG · choosing models & patterns · the NVIDIA NeMo Agent Toolkit · long-running & distributed agents (Google ADK durable sessions, pause/resume, auth.md, A2A fleets) · self-evolving agents (tripartite memory + consolidation) · sovereign AI at the edge (local/$0 inference) · sovereign & self-evolving AI on an NVIDIA DGX (serve/fine-tune/observe/gateway) · vibe-coding & security · the A2A protocol · skill-authoring. See plugins/agentic-coding-fitness/ for details.

📂 Repository Contents

The project is structured week-by-week as our complexity scales up — from a single API call to long-running distributed fleets, self-evolving memory, and sovereign agents running entirely on hardware you own. Each week maps to a plugin skill (above) that recaps it with a $0 lab.

Phase ① Foundation — talk → tools → agents

🔹 Week 2: Claude API Foundations

Talking to modern LLMs programmatically. → skill llm-fundamentals

• week2/claudeapicall.py: basic single-turn API requests · week2/claudestreamingapi.py: streaming tokens · week2/claudemulti_turn.py: conversational state & history · week2/lab/: $0 practice drills.

🔹 Week 3: Tool Use & Smart Assistants

Teaching agents to call external services (function calling). → skill tool-use

• week3/toolsuse.py: function calling (weather, calculator, web search) · week3/buildsmartassistant3tools.py: a full assistant · Tapo Smart Plug Integration (check_tapo.py, scan.py, tapo_config.json): a local HTTP wrapper so Claude controls TP-Link Tapo L530 lights.

🔹 Week 4: Autonomous Pipelines & Hardware

Chaining actions and reaching into physical IoT. → skill agent-loops

• week4/pipeline.py: an autonomous research pipeline (web search → multi-agent synthesis → self-scoring → Markdown reports, with NotebookLM export) · week4/dronecontrol.py: flight patterns on a DJI Tello drone (djitellopy) · week4/openrouterfreemodel.py: a free-model gateway.

🔹 Week 5: The Agent Loop

The reusable REASON → ACT → OBSERVE loop that turns a tool-user into an agent. → skill agent-loops

• week5/autoagent.py: the reusable bounded Agent class (ReAct + stop conditions).

Phase ② Strength — single-agent mastery, reusable tools & knowledge

🔹 Week 6: Full-Stack Agent App (deload / integration)

An agent put behind a real API. → skill vibe-coding-and-security

• week6/src/ (Express/TypeScript/Postgres) · week6/CLAUDE.md + AGENTS.md: context engineering in practice.

🔹 Week 7: MCP & Skills

Reusable tools (MCP) and reusable know-how (Skills). → skill mcp-and-skills

• week7/mcpserver.py, week7/mcpfilesystem.py: MCP servers · week7/agent.py, week7/agenttooldt.py: an MCP client agent · week7/skill.md: a worked Skill.

🔹 Week 8: RAG — Knowledge Agents

Ground answers in your own documents (and prove it with RAGAS). → skill rag-knowledge-agents

• week8/Week8_RAG_Knowledge_Agents_Lab.pdf: the RAG lab.

Phase ③ Endurance — systems that run reliably

🔹 Week 9: Multi-Agent Systems

Sequential / router / parallel-swarm orchestration across frameworks. → skill multi-agent-systems

• week9/ex1_crewai_sequential.py (CrewAI) · week9/ex2_LangGraphSupportGraph.py (LangGraph router) · week9/ex3_ParallelSwarm.py (asyncio swarm) · plus AG2/Anthropic comparisons and 3 workshop PDFs.

🔹 Week 10: Production & Observability

Make a prototype something you can see, stop, and afford. → skills production-and-observability, agent-evaluation

• week10/notebooks/01_hello_graph.py → 05_hybrid_sdk.py: a support-routing system gaining a supervisor, SqliteSaver checkpointing + HITL interrupt(), LangSmith tracing, then a Claude Agent SDK hybrid · week10/GUIDE.md, solutions/.

🔹 Week 11: Mastery — Models & Patterns

Pick the right model, framework, and pattern; the 12-pattern taxonomy. → skills models-and-patterns, agent-drills

• week11/index.html: model wizard + pattern playground + quiz · week11/exercises/: 14 graded MAS drills (ex01–ex14, Beginner → Expert).

Phase ④ Performance — memory, GraphRAG, production frameworks, fleets

🔹 Week 14: Agent Memory with Knowledge Graphs

Durable memory agents remember across runs (Neo4j + GraphRAG). → skill agent-memory-graphs

• week14/agent_memory.py, week14/hotel_kg_builder.py, week14/lab1_hotel_mas.py · week14/NEO4J_TUTORIAL.md · week14/pi-structured-extraction/: a structured-extraction sub-project.

🔹 Week 15: Production GraphRAG

Cypher + GDS, ingestion, GraphRAG across 7 frameworks, and evaluating it. → skills knowledge-graph-mastery, agent-evaluation

• week15/kg_mastery/: the 6-part code companion (fundamentals → building → GraphRAG → evaluation/RAGAS+CI → use cases → reference) · week15/smart_hotel_mas/: a 5-agent CrewAI system over a 4-layer memory stack.

🔹 Week 16: Production Frameworks — NVIDIA NeMo Agent Toolkit

Config-driven multi-agent: register tools, compose YAML workflows, observe. → skill nemo-agent-toolkit

• week16/adding_tools_to_agents.ipynb: tool registration + LlamaIndex RAG tool · week16/multi_agent_orchestration.ipynb: supervisor → specialists with HITL.

🔹 Week 17: Long-Running & Distributed Agents (Google ADK + A2A)

Agents that pause for days and resume without losing context, and delegate across services. → skills long-running-and-distributed-agents, a2a-protocol

• week17/checkpoints/checkpoint1_state_machine.py → checkpoint6_fleet.py: 6 offline steps (durable state → restart-survival → webhook resume → sub-agents → A2A cards → fleet capstone) · week17/hr_onboarding/: a live ADK onboarding agent · week17/authmd_adk/: auth.md × ADK — store the durable grant, re-mint a scoped token at every wake.

Phase ⑤ Sovereignty & Self-Improvement — the stack you own, that gets better

🔹 Week 18: Production Loops, Self-Evolving Memory & Sovereign Edge AI

Three interactive web apps + runnable demos that take the agent stack to production, make it learn, and take it off the cloud. → skills agent-loops (extended), self-evolving-agents, sovereign-ai-edge

• week18/agent_loop/ — the loop as a production discipline via the Claude Agent SDK: built-in & custom tools, PreToolUse/PostToolUse safety hooks, resumable sessions, subagent orchestration, and max_turns/max_budget_usd caps. A clickable streaming web app (tutorial_server.py, port 8090) + 9 demos (step01_hello_agent → step09_production). Uses your claude CLI sign-in — no API key.
• week18/self_evolving_agent/ — turn a stateless agent into one that remembers, learns, and gets cheaper via the Tripartite Memory Model (episodic SessionDB + semantic MEMORY.md/USER.md + procedural SKILL.md library) and a background consolidation loop → compound returns (~64% fewer turns / ~66% lower cost by run 5). Live visualizer (port 8088) + step-by-step guide (port 8090); 7 checkpoints (1–6 offline, $0).
• week18/sovereign_ai_edge/ — run the whole stack on hardware you own with zero cloud dependency and $0 per token: local OpenAI-compatible inference (Ollama), RAM-based hardware sizing, quantization math, LoRA/NeMo fine-tuning, on-device tool-calling agents, a Smart-Hotel HVAC demo, and a live air-gap sovereignty audit. Web app on port 8091 + 9 demos.
• Comprehensive write-ups per folder (README.md/TUTORIAL.md) plus tutorial PDFs: agent_loop_comprehensive_tutorial.pdf, self_evolving_agent_tutorial.pdf, sovereign_ai_edge_tutorial.pdf.

🔹 Week 19: Sovereign & Self-Evolving AI on a DGX

Five interactive web apps that take the whole stack onto an NVIDIA DGX — run/serve, fine-tune, observe, self-evolve, and gateway — grounded in NVIDIA's dgx-spark-playbooks. Every app runs REAL (a live Ollama/vLLM/DGX endpoint) or SIM (a faithful simulator — no GPU needed); cloud cost always $0.

• 👉 Start here: the step-by-step walkthrough → week19/README.md — walks you through all five apps in order, chapter by chapter.
• week19/sovereign_dgx/ (port 8092) — run + serve + manage models on a DGX: Ollama, vLLM, llama.cpp, TensorRT-LLM, NVFP4 quantization, multi-Spark scale-out, air-gap audit.
• week19/dgx_finetune/ (port 8093) — adapt a model to your domain: LoRA/QLoRA with NeMo AutoModel + Unsloth, dataset prep, training loop, eval, GGUF/NVFP4 export.
• week19/dgx_observability/ (port 8094) — see, measure, judge a sovereign agent: OpenTelemetry tracing → Arize Phoenix, metrics, LLM-as-judge evals, + a NeMo Agent Toolkit workflow.
• week19/self_evolving_agent_v2/ (port 8095) — the Week 18 self-evolving agent, made sovereign: a switchable brain (DGX ↔ Claude) + tripartite memory on the DGX that learns over time.
• week19/dgx_litellm/ (port 8096) — the serving gateway: one OpenAI URL over all backends with LiteLLM — routing, fallbacks, hot-swap, virtual keys/budgets, logging → Phoenix.

---

🛠️ Getting Started

1. Requirements

Ensure you have Python 3.10+ installed on your machine.

Clone the repository and set up a virtual environment:

git clone https://github.com/your-username/AgenticCoding.git
cd AgenticCoding
python3 -m venv venv
source venv/bin/activate  # On Windows use: venv\Scripts\activate

2. Install Dependencies

pip install -r requirements.txt

3. Environment Variables

To authenticate with the models, create a .env file in the root directory:

ANTHROPIC_API_KEY="sk-ant-api03-YourAnthropicKeyHere..."

4. Hardware Configuration (Optional)

• Tapo Lights: Edit tapo_config.json with your TP-Link account credentials and local IP address of your light bulb.
• Tello Drone: Connect your computer directly to the Tello's Wi-Fi network before running week4/dronecontrol.py.

---

🎯 Who is this for?

• Developers & Programmers looking to elevate their workflow with AI.
• Tech, Startup, and Product Innovators.
• Anyone with basic coding knowledge ready to embrace the future of AI-native, Agent-based development.

🌟 Our Goal

• Build Real Stuff
• Solve Real Problems
• Generate Real Impact

Come join us every Tuesday, stretch those brain muscles, and let's craft the future of Agentic AI together! 💪🤖