ai-agent-lab

AI Agent Systems Lab

A hands-on learning environment for modern AI agent engineering using Claude Code's native runtime — skills, hooks, sub-agents, MCP servers, agent teams, and context engineering.

Built to practice the patterns that are replacing static LangChain-style prompt pipelines with runtime-native orchestration.

What's Inside

Skills (Reusable Agent Workflows)

Hooks (Deterministic Lifecycle Guardrails)

MCP Server (Custom Tool Extension)

mcp-servers/task-tracker/ — A production MCP server built with the official Python SDK (FastMCP):

• 6 tools: add, list, complete, delete, bulk_add, get_summary
• Resources: lazy-loaded task data
• Prompts: standup report generator, prioritization helper
• Full integration test harness

Slash Commands (Project Workflows)

Agent Definitions

Documented patterns for building agent teams:

• Researcher — Deep codebase exploration without context pollution
• Planner — Concrete implementation plans from synthesized research
• Executor — Isolated code changes (worktree support)
• Validator — Independent verification (fresh eyes, no confirmation bias)
• Coordinator — Runtime orchestration pattern (the parent context)

Architecture

┌── Claude Code Session (Coordinator) ─────────────────────────┐
│                                                               │
│  Context Window                                               │
│  ├── System prompt (~4K tokens)                              │
│  ├── CLAUDE.md (project rules, always loaded)                │
│  ├── Memory index (cross-session knowledge)                  │
│  ├── Skill descriptions (lazy routing metadata)              │
│  └── Conversation + tool results (growing, compacted)        │
│                                                               │
│  Orchestration: runtime decisions between phases              │
│  ├── Spawn sub-agents (isolated context windows)             │
│  ├── Fan-out parallel research                               │
│  ├── Synthesize findings (never delegate understanding)      │
│  └── Checkpoint with user at critical junctures              │
│                                                               │
│  External:                                                    │
│  ├── Hooks (deterministic, zero token cost, shell scripts)   │
│  ├── MCP servers (tool extension via stdio protocol)         │
│  └── Files (persistence beyond context compaction)           │
└───────────────────────────────────────────────────────────────┘

Key Concepts

Why Runtime Orchestration > Static Pipelines

# OLD: Static pipeline (LangChain-style)
chain = research | plan | execute | validate  # same path every time
chain.invoke("add a semicolon")               # overkill
chain.invoke("rewrite auth system")           # underkill

# NEW: Runtime orchestration (agent-native)
# Coordinator classifies, adapts, skips, retries based on actual results
# "Add semicolon" → direct edit (500 tokens)
# "Rewrite auth" → full team pipeline (3000 tokens, correct result)

Context Engineering Principles

• CLAUDE.md = always loaded → keep it SHORT (build commands + critical rules only)
• Memory = cross-session knowledge → index is cheap, content loads on demand
• Skills = lazy-loaded workflows → descriptions route, full instructions load on invoke
• Sub-agents = context isolation → intermediate results stay in child, parent gets summary
• Hooks = zero context cost → deterministic enforcement outside the LLM

Token Budget Mental Model

Getting Started

Use the lab

cd ~/Documents/GitHub/agent-lab
# Claude Code auto-loads .claude/settings.json, CLAUDE.md, .mcp.json
# All hooks, permissions, and MCP servers activate automatically

Try the commands

/project:research MCP protocol architecture
/project:explain context compaction
/project:smart-do add a health check endpoint
/project:agent-team implement user preferences feature
/project:context-audit

Connect the MCP server

The task-tracker MCP server connects automatically when you start Claude Code in this directory. Approve it when prompted.

# Test it manually:
cd mcp-servers/task-tracker
uv run python test_server.py

Project Structure

agent-lab/
├── CLAUDE.md                    ← Project conventions (always loaded)
├── .mcp.json                    ← MCP server configurations
├── .claude/
│   ├── settings.json            ← Hooks, permissions, project config
│   └── commands/                ← Slash commands (/project:*)
├── skills/                      ← Reusable skill definitions (SKILL.md)
├── agents/                      ← Agent architecture documentation
├── hooks/                       ← Shell scripts for lifecycle events
├── mcp-servers/
│   └── task-tracker/            ← Working MCP server (Python/FastMCP)
├── experiments/
│   ├── 01-context-engineering/  ← How the context window works
│   ├── 02-skill-system/         ← Skill loading & routing
│   ├── 03-hooks/                ← All hook types & lifecycle events
│   ├── 05-memory-engineering/   ← Memory architecture & examples
│   └── 06-runtime-orchestration/← Runtime vs static pipelines
└── docs/
    ├── architecture.md          ← Master architecture reference
    └── quick-reference.md       ← Decision cheatsheet

Learning Path

1. Context Engineering → experiments/01-context-engineering/
2. Sub-Agents → agents/researcher.md, agents/coordinator.md
3. Skills → experiments/02-skill-system/, skills/*/SKILL.md
4. Hooks → experiments/03-hooks/HOOK_TYPES.md, hooks/*.sh
5. MCP → experiments/04-mcp-architecture/, mcp-servers/task-tracker/
6. Memory → experiments/05-memory-engineering/
7. Runtime Orchestration → experiments/06-runtime-orchestration/

Resources

Core Documentation:

• Claude Code Docs
• Claude Code Best Practices
• Claude Agent SDK
• MCP Specification

Agent Harness Comparison:

• Three Kingdoms of CLI Agents — Claude Code vs OpenCode vs Pi philosophy comparison (read first)
• OpenCode (sst) — configurable agent product
• Pi (Mario Zechner) — minimal programmable harness

Deep Dives:

• Skills Complete Guide (PDF)
• Hooks Mastery
• Sub-agents & Skills Guide
• Agent Teams Guide
• Agent Loops & Tool Calls
• Build Custom Agent Framework
• Claude Code Skills, Subagents, Hooks Overview

License

MIT