Abhijoy Sarkar 推出的这门 1.5 小时中级 Python 课程，旨在不依赖 LangChain 等第三方框架的情况下，通过纯手写约 1300 行代码，深入剖析 AI 智能体的核心逻辑。课程涵盖 ReAct 循环、原生工具调用、三大内存系统及多智能体协作，帮助具备基础 Python 知识的学习者实现全栈自研。
Published 6/2026
Created by Abhijoy Sarkar
MP4 | Video: h264, 1920×1080 | Audio: AAC, 44.1 KHz, 2 Ch
Level: Intermediate | Genre: eLearning | Language: English | Duration: 49 Lectures ( 1h 41m ) | Size: 805.8 MB
Build your own ~1,300-line agent framework in pure Python — tools, memory, planning, multi-agent, evals. No black boxes.
What you’ll learn
⚡ Build a reusable agent loop — observe, decide, act, feed back, stop — with a correct stopping condition
⚡ Give an agent tools from scratch: parse tool calls, dispatch them, and handle a hallucinated tool without crashing
⚡ Implement working, episodic, and semantic memory — and write cosine similarity and top-k retrieval yourself from scratch
⚡ Make an agent reason with ReAct, reflection, and tree-of-thoughts, and route each task to the right strategy
⚡ Coordinate multiple agents with a delegation cap, evaluate them (success rate, cost, regressions), and harden for production with caps, logging, and guardrails
Requirements
❗ Solid Python: functions, classes, and dicts. You’ve called an LLM API at least once.
❗ An API key for any OpenAI-compatible or Anthropic model (a few dollars of usage covers the labs).
❗ No machine-learning math required — the one bit (a dot product) is taught from scratch. No LangChain, no vector DB.
Description
This course contains the use of artificial intelligence.
An AI agent is just an LLM in a loop that chooses its own next action. Once you can write that loop — about twenty lines of plain Python — every “agent framework” on the market stops being magic.
This course is the un-magic. You build your own agent framework from scratch, one concept at a time, and walk away understanding every line of it — and able to read LangChain, CrewAI, or AutoGen like an open book.
What you’ll build, across nine short modules: the reusable agent loop with a correct stopping condition; tool use from scratch (parse, a registry and dispatch, and graceful handling when the model invents a tool); three kinds of memory — working, episodic, and semantic retrieval you build with your own cosine similarity in numpy; planning with ReAct, reflection-and-retry, and a real beam-search tree-of-thoughts; multi-agent systems with a coordinator and a delegation cap so it can never loop forever; evaluation — the part almost no course teaches — a harness, the three metrics that matter, a failure taxonomy, and regression diffs; production hardening with cost and step caps, structured logging, and tool guardrails; and a capstone that consolidates it into one clean API and builds a flagship agent end to end.
Honest about scope: “from scratch” means you build everything that turns a model into an agent. The one black box we keep is the language model and its embeddings, called through a thin wrapper. It’s about 1,300 lines of pure Python — numpy and a provider client only. No LangChain, no vector database.
You need solid Python and to have called an LLM API once. You do not need any machine-learning math — the one bit, a dot product, is taught from scratch. By the end you won’t just have working code; you’ll have the judgment to design agents that don’t burn money or fail silently, and you’ll understand what every framework on the market is really doing.
Who this course is for
Working developers who use LLMs but treat agents as magic and want first-principles mastery.
Engineers who’ve wired up LangChain or CrewAI without understanding what’s underneath.
Builders who want to ship reliable agents — with evals, caps, and guardrails — not just demos.
Password/解压密码www.tbtos.com
转载请注明：0daytown » Build an AI Agent From Scratch in Python (No LangChain)