Build Your Own Agent — AI skill for Claude Code

Build Your Own Agent

**[中文版 →](README_zh.md)**

**What I cannot create, I do not understand.** — Richard Feynman

4 production Agent projects. ~600,000 lines of code. Distilled into 11 architecture modules and 6 emergent behavior patterns.

This is a step-by-step guide to building AI Agents from scratch — not from theory, not from framework docs, but from patterns extracted from real, working code.

Architecture Overview

flowchart TB
    subgraph SE["🔒 Security / Policy Envelope — optional, Agent-unaware"]
        direction TB
        IF["📡 Interface
CLI · Chat Platforms · Web · API inbound · MCP inbound · ACP"]

        subgraph GR["🚪 Gateway + Routing — optional, multi-user/multi-agent only"]
            G["Gateway
Auth · Rate Limit
Method Dispatch"]
            R["Routing
Which Agent?
Which Session?
Binding Match"]
        end

        subgraph AC["⭐ Agent Core — the Orchestrator"]
            direction TB
            CORE["Conversation Loop · Context Construction · Sub-Agent Delegation · Iteration Budget
↕ Session Store (SQLite/JSONL) · Config Store (YAML/JSON)"]

            subgraph PEERS[" "]
                direction LR
                M["🧠 Model
Provider Routing
Failover Chain
Prompt Cache
Token Estimation
↕ No storage"]
                T["🔧 Tools
Register · Discover
Dispatch · Parallel
MCP outbound
Security Approval
↕ Skill Store"]
                MEM["🧩 Memory
Dual Role:
① Tool — LLM calls
② Infra — Core injects
↕ Memory Store"]
                CTX["📐 Context Mgr
Compression
Tool Result Pruning
Token Budget
↕ No storage"]
            end
        end
    end

    RT["🖥️ Runtime — local process / Docker / OpenShell sandbox / cloud"]
    INFRA["🏗️ Infrastructure — Blueprint plan/apply · Health check · Rollback"]

    IF --> GR
    GR --> AC
    AC ~