FAQ
Error quick reference
Look it upWhat I'm seeing is…
Click one to see.
What do I need to learn for Agent development?
Follow this course's order: a little Python → function calling → the agent loop → tool design → memory → skills and MCP → frameworks, security, evaluation. The first five parts don't require any framework; a single openai library is enough.
Which language should I use for Agent development?
Python and TypeScript are the mainstream choices. The official MCP SDK, OpenAI Agents SDK, Claude Agent SDK, and LangGraph all have Python versions; the Claude Agent SDK and MCP also have TypeScript versions. The principles are language-agnostic, and the loop in this course can be written in any language.
What's the difference between Agent development and "AI application development"?
AI application development is the broader category: calling a model once for translation or customer service also counts. Agent development specifically means letting the model decide, within a loop, which tools to call and how many steps to take to complete a task. Many applications actually only need a single call or a fixed workflow, not necessarily an agent—see the first section.
Can the Claude Agent SDK be used with domestic (Chinese) large models?
The official documentation we found only describes using an Anthropic API Key, or using Claude through platforms such as Amazon Bedrock, Google Cloud, and Microsoft Foundry; it does not state that models from other vendors can be connected. If you want to use domestic or local models, you can write it by hand (this course) or use the OpenAI Agents SDK—see how to choose a framework.
Are local models good enough?
All the code in this course has been run successfully on a local qwen3:8b, including reading files, using skills, and connecting to MCP. Smaller models make mistakes more easily (for example, writing wrong paths), so feeding errors back to the model and requiring human confirmation for dangerous operations becomes even more important. For complex tasks, you can develop and debug with a local model first, then switch to a cloud large model.
What's the difference between MCP and Function Calling?
See the dedicated section. In one sentence: Function Calling is the format between the model and your program, while MCP is the protocol between your program and tool servers; the two are used together.
Where can I download the code?
myagent.zip (the complete agent, skills, MCP server, and example work records), plus the standalone step1_call.py, step2_loop.py, and step_sdk.py.
Versions used in our testing: openai 3.57.0, mcp 2.3.0, openai-agents 0.25.0, Ollama 0.40.16, qwen3:8b.