纯Python从零构建AI智能体：部署AWS Bedrock实战教程_CG群

本教程由Prasad Yarra更新，教你用纯Python无框架构建AI智能体，实现感知-思考-行动循环、ReAct推理、工具调用与内存系统，并集成Gemini API，通过Terraform部署到AWS Bedrock AgentCore。

原始标题：Python for AI Agents: Build And Deploy Agent From Scratch

Python for AI Agents: Build And Deploy Agent From Scratch

这门由 Prasad Yarra 于 2026年9月 更新的中级视频教程，核心专注于在不依赖任何外部框架的情况下，纯手写 Python 构建、部署并固化 AI 智能体（Agent）。在 1 小时 54 分钟的实战演练中，学员将从零实现“感知-思考-行动”循环、ReAct 推理模式、工具调用分发器以及防崩溃的长期内存系统，并最终接入 Google Gemini API 驱动智能体。

该课程的最大亮点在于全实战与云端部署，讲师在屏幕上完整展示了从代码编写到报错调试的全过程，并使用 Terraform 将智能体部署至 AWS Bedrock AgentCore。课程还涵盖了应对真实限流（429 错误）的重试逻辑，以及通过 AWS Secrets Manager、IAM 最小权限和 CloudWatch 监控进行生产级安全加固，非常适合想要彻底掌握 AI 智能体底层原理的 Python 开发者。

Perceive-think-act, tools, memory, ReAct, and a real Gemini agent – deployed to AWS and torn down on screen.

What you’ll learn
⚡ Build a working AI agent in plain Python with zero frameworks and zero external dependencies
⚡ Implement the perceive-think-act loop that underlies every agent architecture
⚡ Give an agent real tools using a function-calling pattern with a registry and dispatcher
⚡ Design short-term (context window) and crash-safe, atomic-write long-term memory for an agent
⚡ Implement the ReAct reasoning pattern and multi-step task planning from scratch
⚡ Connect a real LLM (Google Gemini) and let it choose which tools to call
⚡ Deploy a production agent to AWS Bedrock AgentCore using Terraform
⚡ Add retry logic that survives real rate limits, and harden a deployment with Secrets Manager, least-privilege IAM, and CloudWatch monitoring

Requirements
❗ Comfortable with core Python: functions, classes, dictionaries, basic file I/O
❗ No prior AI/ML or cloud experience needed
❗ A computer that can run Python 3.10+ (any OS)
❗ For Module 5 only: a free Google Gemini API key and an AWS account (only needed if you want to follow the live deployment lessons yourself)

Description
This course contains the use of artificial intelligence. All lectures use AI-generated voice narration.

You can build an AI agent in a few lines with a framework. But when it does something unexpected, would you know why?

This course shows you what’s actually going on inside an agent – by building one yourself, from a 30-line rule-based loop all the way to a real Gemini-powered agent deployed on AWS Bedrock AgentCore with Terraform, monitored, and torn down again, on screen.

The screen recordings were captured while writing and running the code, not edited afterward to skip the mistakes. When something breaks — a wrong log group, a retry that never fires, a rate limit hit — you see the bug and the fix, because that’s how you learn what the code does.

You’ll build, in order

✅ A minimal rule-based agent using nothing but the Python standard library

✅ A reusable perceive-think-act loop

✅ A tool-calling system (calculator, clock, text utilities) with a registry and dispatcher

✅ Short-term memory (conversation history / context windows) and durable long-term memory (crash-safe atomic file writes)

✅ A ReAct-style reasoning loop and a multi-step planner, from scratch, no framework

✅ A real agent backed by Google’s Gemini API, choosing which of your tools to call

✅ A production deployment of that agent to AWS Bedrock AgentCore, provisioned with Terraform

✅ Retry logic that survives real rate limits – you’ll fire concurrent requests, hit a real 429, and watch it recover

✅ A hardened, secured version: secrets in AWS Secrets Manager, least-privilege IAM, real CloudWatch monitoring – and a real terraform destroy to close it all out

Along the way you also get

✅ Hands-on exercises for every module, with worked solutions

✅ A short quiz plus a “debug this” challenge after every lesson

✅ A capstone project – a Research Notes Agent combining tools, durable memory, and multi-step Gemini reasoning in one build, with a starter (TODOs included) and a finished solution

If you already know some Python and want to understand how an AI agent actually works by building one yourself, this course is for you.

Who this course is for
⭐ Python developers who want to understand how AI agents work under the hood, not just call a framework
⭐ Developers evaluating agent frameworks (LangChain, Strands, etc.) who want to know what those frameworks are doing for them
⭐ Anyone who wants a real, working example of deploying and hardening an LLM agent in production on AWS

隐藏内容

此处内容需要权限查看