Q&A

These questions came up during the workshop and are useful on the side.

Recording

The workshop recording is linked from Building AI Agents with MCP, PydanticAI and OpenAI.

AI assistance in notebooks

For AI help while working in notebooks, use VS Code or Cursor notebook support, or use Google Colab with Gemini support. The examples use Jupyter directly, but the same notebook can be opened from an IDE.

minsearch and RAG

The workshop uses text search in minsearch, not vector search. That still counts as RAG because we retrieve data from a knowledge base and use it to augment the answer. The retrieval backend can be text search, vector search, a SQL query, or any other function.

Responses API message format

With the Responses API, the model output objects can be appended back to the input list directly. With the older Chat Completions API, the message format is different and role fields are handled differently.

OpenAI API key

Set OPENAI_API_KEY as an environment variable before running the notebook. If you are not sure how to do that on your operating system, ask an assistant for OS-specific shell instructions. You can also pass api_key directly to the client while testing, but do not commit it.

ToyAIKit and production

toyaikit is for teaching, workshops, and quick experiments. For production-grade agent work, use a tested framework such as OpenAI Agents SDK, PydanticAI, LangChain, or your own well-tested loop.

Notebook chat input

The interactive notebook chat uses Python input(). The runner reads a question, stops if the question is stop, and otherwise passes the text to the same agent loop used by runner.loop(...).

uv

uv creates an isolated environment and pins dependencies in uv.lock. That makes the workshop easier to reproduce and avoids packages from other projects interfering with the notebook.

Prompt logic and function logic

Put behavioral instructions in the prompt. Put actions in functions. Searching a database, sending an email, appending a row, and calling an API belong in tools. The docstring is part of the tool description the model reads, so use it to explain when and how the tool should be used.

PydanticAI and multiple providers

PydanticAI can use different model providers through the model string. The example starts with OpenAI, then switches to an Anthropic model with model="anthropic:claude-3-7-sonnet-latest". That requires an Anthropic key in the environment.

Anthropic key

Anthropic API access is paid, similar to OpenAI API access. The workshop does not walk through the full signup flow. The Anthropic example assumes a configured environment variable.

MCP and databases

Postgres is a good example of why MCP exists. If many agents need database access, one team can expose database tools through an MCP server and agent teams can call that server instead of each writing their own database integration.

Hallucination mitigation

To reduce hallucinations, start with evaluations and monitoring. The workshop does not implement that layer directly. It builds the tool foundation that later evaluation work can test.

MCP registries

Cursor has a directory of MCP servers that can be added to projects. Docker is an example of a tool you might add through that kind of registry.

FastMCP name

The name is likely inspired by FastAPI. For the workshop, the important part is that FastMCP gives us a small server framework for exposing tools.

Context7 and DeepWiki

Context7 is a ready-made MCP server for framework documentation. DeepWiki is a related idea for GitHub repositories. Both fit the same pattern: index a source of information and let an agent consult it through a tool.