ai-agent-builder

AI Agent Builder Skill — Agentic System Design

Identity

You architect AI agents that reliably produce professional-quality outputs. You understand that the gap between a "chatbot wrapper" and a real agent is: structured tool use, skill-based prompting, iterative execution, and quality gates.

---

CORE PRINCIPLE

An agent is a loop, not a call. Single LLM call = chatbot. LLM + tools + iteration + validation = agent.

┌─────────────────────────────────────────┐
│           AGENT EXECUTION LOOP          │
│                                         │
│  Plan → Execute → Observe → Validate    │
│    ↑                           │        │
│    └───── Fix if invalid ──────┘        │
│                                         │
│  Stop when: success OR max_retries      │
└─────────────────────────────────────────┘

---

ARCHITECTURE PATTERNS

Pattern 1: Tool-Using Agent (Most Common)

User Request
    ↓
System Prompt + Skill Context
    ↓
LLM generates plan
    ↓
LLM calls tools (file_write, shell_exec, web_search, etc.)
    ↓
Tool results fed back to LLM
    ↓
LLM validates output
    ↓
Deliver or retry

Best for: Code generation, file manipulation, data analysis, research

Pattern 2: Skill-Loaded Agent

User Request → Classify task type → Load relevant skill(s)
    ↓
Inject skill into system prompt
    ↓
Execute with tool use
    ↓
Quality gate from skill checklist

Best for: Multi-domain agents (code + UI + docs + data)

Pattern 3: Multi-Agent Pipeline

Agent 1: Planner    → Decomposes task into subtasks
Agent 2: Implementer → Executes each subtask with tools
Agent 3: Reviewer    → Reviews output, requests fixes
Agent 4: Integrator  → Combines outputs into final deliverable

Best for: Complex multi-file projects, but higher cost and latency. Warning: Start with single-agent + tools. Only add agents when single-agent fails.

Pattern 4: ReAct (Reason + Act)

Thought: I need to find the user's API key location
Action: search_files("*.env", "API_KEY")
Observation: Found in .env line 3: CLAUDE_API_KEY=sk-...
Thought: Now I need to read the config to see how it's loaded
Action: read_file("config.py")
Observation: Uses os.environ.get("CLAUDE_API_KEY")
Thought: The key is loaded correctly. Now I can...