Andrew Ng's Four Agent Patterns, De-Frameworked

What it does This repository implements the four agentic patterns Andrew Ng outlined for DeepLearning.AI—reflection, tool use, planning, and multi-agent collaboration—using nothing but direct API calls to Groq. It packages the code as both step-by-step Jupyter notebooks and a small, installable Python library called agentic-patterns.

The interesting bit The author deliberately avoids the usual orchestration frameworks—no LangChain, no CrewAI, no LlamaIndex—so the logic stays transparent. Each pattern is boiled down to a single, inheritance-based Python class (ReflectionAgent, ToolAgent, ReactAgent, and Crew) that you can read and modify in minutes.

Key highlights

• Implements reflection, tool use, ReAct planning, and multi-agent crew patterns from scratch.
• Zero framework dependencies; talks directly to the Groq API.
• Includes step-by-step notebooks and a reusable agentic-patterns PyPI package.
• Tool system uses a simple @tool decorator to bind Python functions to the LLM.
• ReactAgent extends ToolAgent to add multi-step reasoning without extra abstraction layers.

Caveats

• Hard-wired to Groq; swapping in another provider means editing the source.
• The multi-agent section references CrewAI’s abstractions, so it may borrow concepts from the very frameworks it avoids.
• The README is heavy on quickstart snippets and YouTube links, light on architectural details.

Verdict Good for developers who want to understand agent internals without wading through framework indirection. Skip it if you need multi-provider support or production observability.