Tag
8 items tagged with "ai-agents"
Build a complete AI agent from data ingestion to deployment. Learn RAG, search, function calling, evaluation, and web deployment — all on YOUR GitHub project.
I unpacked the mysterious /home/oai/ directory that ChatGPT can zip on command and found not leaked OpenAI infrastructure, but a sandbox of document, slide, and spreadsheet “skills” that reveals how the new cross-vendor agent skills standard works in practice.
We start from the coding agent from the prerequisite workshop and turn it into a general-purpose coding agent with two reusable behavior layers: skills and slash commands. Skills are loaded by the agent when the user's request matches a skill description. Commands are invoked by the user with a leading slash and rendered into prompts before the agent acts. Links External resources for this workshop: Starting notebook GitHub fetch skill Prototype implementation Prerequisite coding-agent workshop ToyAIKit OpenCode OpenCode skills documentation AgentSkills spec The agent you will build The final workshop system looks like this: flowchart LR USER["User"] RUNNER["ToyAIKit runner<br/>OpenAI Responses"] LLM["OpenAI model"] CODETOOLS["Coding tools<br/>read, write, tree, bash, search"] SKILLTOOL["skill(name) tool"] LOADER["SkillLoader"] SKILLS["skills/*/SKILL.md<br/>scripts and templates"] COMMANDS["commands/*.md"] COMMANDTOOL["execute_command(name, args)"] USER -->|plain request| RUNNER USER -->|/command| RUNNER RUNNER --> LLM RUNNER --> CODETOOLS RUNNER --> SKILLTOOL SKILLTOOL --> LOADER LOADER --> SKILLS RUNNER --> COMMANDTOOL COMMANDTOOL --> COMMANDS The project stays small enough to understand in a notebook, but it mirrors the pieces used by real coding agents. The coding tools let the model read, write, search, and run commands. The skill loader turns SKILL.md files with YAML frontmatter into tool-loadable instructions. The command loader turns markdown files like review.md or test.md into reusable prompt templates. Appendix A file list for the workshop materials is in Appendix: workshop files.
We start with a DataTalks.Club Data Engineering Zoomcamp FAQ assistant, then add checks that keep the agent on topic, block unsafe responses, and show how to cancel wasted work when a guardrail fails. The workshop uses the OpenAI Agents SDK for built-in guardrails, then rebuilds the same idea with tools and plain asyncio so you can use it with other agent frameworks. Links The external resources: Related course: AI Bootcamp: From RAG to Agents FAQ data used by the agent AI Hero email course for the docs.py loader OpenAI Agents SDK guardrails documentation The notebook you will build The final notebook has guardrails around a tool-using FAQ agent: flowchart LR USER["User question"] INPUT["Input topic guardrail"] FAQ["FAQ assistant agent"] SEARCH["search_faq tool<br/>minsearch index"] OUTPUT["Output safety guardrail"] ANSWER["User-facing answer"] OPENAI["OpenAI model calls"] USER --> INPUT INPUT -->|passes| FAQ INPUT -->|trips| ANSWER FAQ --> SEARCH FAQ --> OPENAI INPUT --> OPENAI OUTPUT --> OPENAI FAQ --> OUTPUT OUTPUT -->|passes| ANSWER OUTPUT -->|trips| ANSWER The base agent can already search the FAQ, but it tries to answer unrelated questions too. The input guardrail blocks questions outside the course domain. The output guardrail checks the agent response for policy problems such as promising deadline extensions or writing homework for a student. The later parts show the same checks as tools and as a small async runner that can cancel work when a guardrail trips.
We build a deep research agent over the DataTalks.Club podcast archive. The workshop starts by downloading and indexing YouTube transcripts, then turns that ingestion code into a durable Temporal workflow. After the data is searchable, we build a Pydantic AI research agent, add a summarization sub-agent for long transcripts, and wrap the agent run in Temporal too. Links The main resources for this workshop: Temporal CLI install notes DataTalks.Club podcasts The system you will build The final system looks like this: flowchart LR YT["YouTube transcripts"] CACHE["Cached transcript files<br/>GitHub fallback"] FLOW["flow/<br/>Temporal ingestion"] ES["Elasticsearch<br/>podcasts index"] AGENT["agent/<br/>Pydantic AI research agent"] SUM["Summarization sub-agent"] TW["Temporal workflow<br/>agent run"] OPENAI["OpenAI"] YT -->|fetch_subtitles| FLOW CACHE -->|fetch_transcript_cached| FLOW FLOW -->|index_video| ES AGENT -->|search_videos| ES AGENT -->|summarize| SUM AGENT -->|model call| OPENAI SUM -->|model call| OPENAI TW -->|durable execution| AGENT The ingestion side has the parts that usually fail in production: network calls, YouTube blocking cloud IPs, proxies, Elasticsearch writes, and long loops over many videos. Temporal gives that side retries, observability, and durable execution. The agent side uses the indexed data to answer questions from the podcast archive and then uses Temporal again so long agent runs can survive failures.
We build a course FAQ assistant from the bottom up. First we expose a plain Python search(query) function to the OpenAI Responses API. Then we turn the same idea into a reusable agent loop, compare toyaikit, OpenAI Agents SDK, and PydanticAI, and finally move the FAQ tools behind an MCP server that can be used from a notebook, PydanticAI, Cursor, and VS Code. Links The main resources: AI Bootcamp: From RAG to Agents Prerequisite workshop: Building a Coding Agent Data Engineering Zoomcamp FAQ source document Parsed FAQ JSON FAQ parsing notebook The system you will build The final setup looks like this: flowchart LR NOTEBOOK["Jupyter notebook"] OPENAI["OpenAI Responses API"] FRAMEWORKS["Agents SDK<br/>PydanticAI"] MCPCLIENT["MCP clients<br/>toyaikit, PydanticAI, Cursor"] MCPSERVER["FastMCP server<br/>SSE or stdio"] TOOLS["FAQ tools<br/>search, add_entry"] INDEX["minsearch index<br/>FAQ JSON"] NOTEBOOK -->|function calling| OPENAI NOTEBOOK --> FRAMEWORKS FRAMEWORKS -->|tool calls| TOOLS NOTEBOOK -->|MCP client| MCPCLIENT MCPCLIENT -->|MCP protocol| MCPSERVER MCPSERVER --> TOOLS TOOLS --> INDEX The FAQ data comes from the Data Engineering Zoomcamp FAQ. The first half of the workshop keeps the tools inside the notebook so you can see the agent loop directly. The second half moves the same tools into mcp_faq/, which makes them reusable by any MCP client.
We build a small project bootstrapper for Django: a coding agent that takes a plain-language app request, copies a working Django template, reads and writes files through tools, and iterates until the generated app runs. The first implementation uses the OpenAI Responses API through ToyAIKit, then we try the same idea with OpenAI Agents SDK, PydanticAI, Anthropic, and Z.AI. Links The main resources: ToyAIKit Django template repo Todo app made with Z.AI Related course: AI Bootcamp: From RAG to Agents Related workshop: Hands-on with AI Agents and MCP The app you will build The coding agent is a notebook-based chat interface backed by an LLM and a small set of filesystem tools. You give it a request like to-do list. The agent edits a copied Django template and leaves you with a project you can run. flowchart LR USER["You<br/>short app request"] CHAT["Jupyter chat UI<br/>ToyAIKit"] RUNNER["Agent runner<br/>Responses API or framework"] TOOLS["AgentTools<br/>read, write, tree, grep, bash"] DJANGO["Copied Django template<br/>project folder"] LLM["LLM provider<br/>OpenAI, Anthropic, Z.AI"] USER -->|type request| CHAT CHAT --> RUNNER RUNNER -->|tool calls| TOOLS TOOLS -->|modify files| DJANGO RUNNER -->|messages and tools| LLM DJANGO -->|make run| USER Two screenshots show what the finished workshop output looks like. The first one shows the notebook chat after the agent plans and starts calling file tools: The second one shows one of the generated Django todo apps: Result The simplest version is intentionally small. It runs in Jupyter, uses local filesystem tools, and edits one copied Django project folder. That is enough to understand how larger coding agents work under the hood: prepare a template, expose the right tools, give the model precise instructions, and iterate on the generated code.
We compare AI coding tools by asking them to help build the same small app: Snake in React. The session is a comparison, not a single production build, so the walkthrough follows the tool categories: chat applications, coding assistants and IDEs, project bootstrappers, and agents. Links The main resources: Lovable Snake preview AI Dev Tools Zoomcamp AI Engineering Buildcamp RAG agents workshop LLM Zoomcamp The comparison The same app appears in several workflows: flowchart LR PROMPT["Prompt<br/>implement snake in react"] CHAT["Chat apps<br/>ChatGPT, Claude, DeepSeek, Ernie, Microsoft Copilot"] IDE["Coding assistants and IDEs<br/>Claude Code, GitHub Copilot, Cursor, Pear"] BOOT["Project bootstrappers<br/>Bolt, Lovable"] AGENTS["Agents<br/>tools, file access, computer use"] CHATAPP["Local app<br/>snake-chatgpt"] CLAUDEAPP["Local app<br/>snake-claude-code"] PREVIEW["Hosted app<br/>Lovable preview"] PROMPT --> CHAT CHAT --> CHATAPP IDE --> CLAUDEAPP BOOT --> PREVIEW AGENTS --> IDE AGENTS --> BOOT The workshop code includes two Vite React projects. snake-chatgpt keeps the game in one component and is useful for seeing what a chat app can produce from a short prompt. snake-claude-code splits the page shell from the game component, adds score, start and restart flow, WASD controls, and a walls versus pass-through mode.