Building an AI Agent from Scratch: From 50 Lines of Python to MCP
1. An agent centers on a loop that reassesses tool results
An agent is a program that repeats reasoning → tool execution → observation → deciding what to do next.
- The model uses conversation history to answer or select the tools it needs.
- The program actually executes the tools, and their results are added to the conversation history for the next decision.
- In the roughly 50-line Python example, the program executes tools when `tool_calls` is present; otherwise, it returns the response and stops.
- Conversation history serves as short-term memory within a task.
- A “deterministic loop” describes the code’s control flow. It does not guarantee the model’s choices or task success.
2. Dividing work between local and cloud models
You can keep the same loop while changing where the model runs and which tasks it handles.
- The example starts with the OpenAI API, then connects a local model through Ollama’s compatible API.
- In the author’s experiment, some models returned only statements about using tools instead of structured tool calls, causing the loop to end immediately.
- The ability to produce natural responses and the ability to return executable tool calls need to be checked separately.
- In mixed mode, the local model handles the basic loop and simple tasks.
- Complex questions are delegated to a cloud model through a tool such as `ask_cloud_expert`.
- This setup separates the points where external calls are needed instead of handling every step in the cloud.
3. Extending tools and connecting them through MCP
The agent’s basic loop stays in place as its available capabilities grow.
- Tools expand from checking the date and calculating to web search and reading and writing files.
- The article’s weather and search code returns sample responses, and file operations are simplified, so these should be distinguished from finished capabilities.
- MCP extends connectivity from registering tools directly in code to discovering and invoking them through external servers.
- Client: Retrieves the server’s list of tools and calls the ones it needs.
- Server: Exposes functions you build so other agents can use them too.
- What changes is where tools live and how they are connected and maintained. The loop that returns execution results to the model remains in place.
4. Understanding the minimal implementation reveals what frameworks do
The short example exposes the mechanics, while production services add operational capabilities on top.
- The comparison with Claude Code examines features such as retries, approval before execution, memory across sessions, context compression, and subagents.
- Frameworks also address different problems.
- LangGraph is introduced as an example for arranging state flows, checkpoints, and human approval steps.
- CrewAI and AutoGen illustrate collaboration among multiple agents with different roles.
- Understanding a small loop yourself makes it easier to decide what to implement directly and which capabilities to leave to frameworks or existing products.
