GeekCoding101 make your way to geek

Agent Stack Insider: Hands-On Guides to AI Agents, MCP and Agentic Frameworks

7 posts on agentstackinsider, newest first.

7posts
2025first post

Agent Stack Insider is GeekCoding101’s hands-on series on the AI agent stack. It covers what an agent actually is, how it calls tools over MCP, which agentic framework to build on, and what happens when you run real agents on real work. If you’re new, start with the two Python build-alongs. If you’re choosing a stack, jump to the framework comparison.

Every post here is written from running the thing on a real machine. There are install steps, real commands, and the errors hit along the way.

Where to start

If you want to…Read
Understand what makes an agent different from a chatbotWhat Is an AI Agent? How to Build One from Scratch in Python: the agent loop explained, plus a working Python agent with tools on the Claude API
Give an agent tools, resources and promptsWhat Is an MCP Server? How to Build One in Python: build, test and connect an MCP server with the Python SDK 2.x
Choose a frameworkAgentic Frameworks Compared: LangGraph, CrewAI & More: LangGraph, LlamaIndex, OpenAI Agents SDK, Google ADK and CrewAI, plus agentic AI coding frameworks like Claude Code
See a coding agent work on a real taskI Ran DeepSeek’s dsh-superagent End to End: installed on a Mac, fixing a real GitHub issue, with its two-layer authorization gate and Mandates explained
Use MCP for security workGhidra MCP: Let AI Help You Reverse Engineer Binaries: Ghidra 12.1.4 and GhidraMCP 2.0 on an Apple Silicon Mac, finding zlib’s Adler-32 in a stripped binary
Self-host an agent platformUnderstanding OpenDAL Storage in Dify: OPENDAL_FS_ROOT, .env settings, Docker volume mounts and permission-error fixes

The agent stack, layer by layer

An “AI agent” is several layers working together. Each post in this series sits on one of them:

  1. The model. The LLM decides what to do next.
  2. The agent loop. Code calls the model, runs the tool it asks for, feeds the result back, and repeats until the task is done. The from-scratch Python post builds this by hand.
  3. Tools and context (MCP). The Model Context Protocol is a standard way to expose tools, resources and prompts to a model. The MCP server post builds one, and the Ghidra MCP post puts one to work on reverse engineering.
  4. The framework. Libraries such as LangGraph, CrewAI, LlamaIndex, OpenAI Agents SDK and Google ADK give you the loop, state and multi-step logic so you don’t write them yourself.
  5. Permissions and control. These rules decide what an agent is allowed to do on your machine or repo. dsh-superagent’s two-layer authorization gate is a concrete example.
  6. The platform and storage. This is where a self-hosted setup keeps its files and data, as in Dify with OpenDAL.

Agent framework vs orchestration tool vs control plane

  • Agent framework: a library you write your agent in. It handles the loop, tool calls and state.
  • Orchestration tool: coordinates several agents or steps into a workflow, deciding who runs when and with which inputs.
  • Control plane: governs agents once they run, covering permissions, policies and visibility into what they did.

The terms often overlap, and many frameworks include some orchestration. For a side-by-side look at real options, see the framework comparison post.

FAQ

What is an agentic framework? It is a library that gives you the building blocks of an AI agent, so you don’t write them from scratch. These include the model–tool loop, state, memory and multi-step control flow. LangGraph and CrewAI are two examples compared in this series.

Do I need a framework to build an AI agent? No. The “What Is an AI Agent?” post builds a working agent with tools in plain Python on the Claude API. A framework becomes useful once you need state, branching workflows or several agents.

What is an MCP server, and why does it matter for agents? An MCP server exposes tools, resources and prompts through the Model Context Protocol. Any MCP-capable client can then use them. The series shows how to build one in Python and how GhidraMCP uses MCP for binary analysis.

What do I need to follow along? The Python posts need Python and, for the agent build, Claude API access. The dsh-superagent and GhidraMCP walkthroughs were done on a Mac, and GhidraMCP specifically on Apple Silicon. The Dify post uses Docker.

Is it free? Yes. Every post is free to read, with no sign-up.