Hello Agent System · AI Nav City

Core facts

Read by a script from the official API, not rewritten by a model.

Repository: https://github.com/heaven999b/hello-agent-system
License: MIT
Languages: Python
Maintainer: heaven999b
Verified: 2026-11-03

What it is

Hello Agent System is an enterprise-level Agent system design training camp. It has 32 lessons, is bilingual in Chinese and English, and each lesson has exercises with tests. The project does not depend on an Agent framework and hand-writes each layer of an enterprise-level Agent from scratch. The project's architecture design was assisted by Claude Opus 5.5.

Main features

The project is async from lesson 02 onward.
From lesson 12 onward it is real multi-process.
The whole project has only one async implementation, used from lesson 02 to lesson 31.
Each layer is hand-written from scratch, with a core of about 3800 lines and multi-process modules of about 2000 lines.
Each handout and document has an equivalent English version.
Messages are dicts in the OpenAI format.

Getting started

First edit the .env file and fill in any OpenAI-compatible interface; the three assignment lines in the example are the content to put into .env, not commands to run in a terminal, and the environment check and demo come after that.
Each lesson follows this flow: read the handout, run the demo, write the exercise, turn the tests green, and go through the self-test checklist.
The author develops and verifies locally with the cliproxyapi gateway and gpt-5.5.

From the README · bash

git clone https://github.com/heaven999b/hello-agent-system.git
cd hello-agent-system
make setup          # 创建 .venv 并安装（核心依赖只有 openai + pydantic）

From the README · bash

LLM_BASE_URL=https://api.openai.com/v1     # 或 DeepSeek / 通义 / 本地 vLLM / 模型网关
LLM_API_KEY=sk-xxx
LLM_MODEL=gpt-4o-mini                       # 需要支持 function calling

From the README · bash

make check-env                                  # 检查模型连通性与工具调用能力
.venv/bin/python lessons/00_overview/demo.py    # 看一眼你将要造出的东西

Who it is for

The course requires being able to read Python functions and classes and having called an LLM API once. Code comments and demo output are currently in Chinese. The agentkit core is an async runtime with real concurrency, real cancellation, bulkheads and streaming, but its state is by default in memory or local files.

In the agentkit core, injection detection and PII redaction are regular expressions. agentkit.distributed lets agentkit run as multiple worker processes on one machine, with crash takeover. agentkit.distributed is based on SQLite and works on one machine only.

Multiple machines require switching to agentkit.contrib, which holds Postgres, Redis, Temporal, OpenTelemetry, LiteLLM and Cedar adapters. The agentkit layers are reference implementations, not hosted products. The Docker / Kubernetes configuration has not actually been started on the local machine.

Redis failover has not been tested. The project is licensed under MIT. ITBuddy is an enterprise IT service desk Agent.

The failure-mode document lists 79 failure modes met in production, giving symptom, root cause, detection and fix. The design review checklist has 199 pre-launch check items graded P0/P1/P2.

Related tools listed on this site.