build-your-own-agents-skill

# Build Your Own Agents

Build AI agents with the same architecture patterns used by Claude Code — without a framework and without reverse-engineering it yourself.

This repo gives you the production checklist, 6 worked examples, and a runnable Python agent you can clone and adapt to your domain in 30 minutes.

## The Problem This Solves

When you ask an AI coding agent to "build me an agent," you get a script that calls an API once and prints the answer. That is not an agent. It is a wrapper.

Production agents like Claude Code use a fundamentally different architecture: a controller loop, tool contracts with schemas, runtime permission enforcement, session state with compaction, hook pipelines for safety, observability, and evaluation.

You should not have to study a production codebase to learn these patterns. **This skill packages them so your coding agent produces production-grade designs automatically.**

## What Changes When You Install This Skill

| Without this skill | With this skill installed |
|---|---|
| One-shot script | Loop-based controller that iterates until the task is done |
| No tool boundaries | Explicit tool contracts with schemas, permissions, and timeouts |
| No safety | Runtime permission enforcement (auto-allow / ask-first / deny) — the same tiered model used by Claude Code |
| No approval flow | Human-in-the-loop approval before high-risk actions |
| No memory | Structured session state with persistence and compaction |
| No logging | Structured JSON observability for every tool call and turn |
| No evaluation | Quality rubrics, safety metrics, and testing cadence |
| No failure handling | Classified failures with bounded retries and graceful degradation |
| "Ship it" | Rollout phases from prototype to production |

## Clone And Run In 5 Minutes

Requires **Python 3.10+** and an [Anthropic API key](https://console.anthropic.com/).

```bash
git clone https://github.com/xuanhieu2611/build-your-own-agents-skill.git
cd build-your-own-agents-s