Agentic Design Patterns — Gemini CLI Extension

A Gemini CLI Extension packaging 28 agentic design pattern skills. Install with a single command and start building production-ready AI agents immediately.

Install

gemini extensions install https://github.com/hajekim/agentic-design-patterns-extension

After installation, restart Gemini CLI. The 28 skills activate automatically when you describe what you want to build.

Current version: v2.2.4 — See CHANGELOG.md for full version history.

What This Extension Provides

28 agent skills across four categories:

Category	Skills
Core Patterns	Prompt Chaining, Routing, Parallelization, Reflection, Tool Use, Planning, Multi-Agent Collaboration
State Management	Memory Management, Learning & Adaptation, MCP, Goal Setting
Reliability	Exception Handling, Human-in-the-Loop, RAG
Advanced Patterns	A2A, Resource-Aware, Reasoning, Guardrails, Evaluation, Prioritization, Exploration
Appendix	Prompt Engineering, GUI Agents, Agentic Frameworks, AgentSpace, AI CLI, Coding Agents, Reasoning Engines

Skills use the DEFINE → PLAN → ACTION workflow and include implementation examples in Google ADK, LangChain, and LangGraph.

How Skills Activate

Gemini CLI reads the name and description of each skill. When your request matches a skill's description, the model automatically loads the full skill instructions.

1,376 trigger phrases across four languages:

# English
"Build a multi-step agent pipeline"          → Prompt Chaining
"set up MCP server"                          → MCP
"choose agent framework"                     → Agentic Frameworks
"thinking model for complex reasoning"       → Reasoning Engines

# 한국어
"프롬프트 체이닝으로 파이프라인 만들어줘"    → Prompt Chaining
"MCP 구성을 해줘"                           → MCP
"메모리뱅크 만들어줘"                       → Memory Management
"추론 모델 언제 써야 해?"                   → Reasoning Engines

# 日本語
"マルチエージェントを構築したい"             → Multi-Agent Collaboration
"MCPサーバーを設定したい"                   → MCP
"RAGパイプラインを作りたい"                 → RAG

# 中文
"帮我构建多智能体系统"                      → Multi-Agent Collaboration
"配置MCP服务器"                            → MCP
"搭建RAG知识库问答系统"                    → RAG

Commands & Agents

Slash Commands

# Browse all 28 patterns grouped by category
/pattern-summary

# Filter by category: core / state / reliability / advanced / appendix
/pattern-summary reliability

# Look up a specific pattern
/pattern-summary planning

# Generate a Python code skeleton for a pattern
/gen-skeleton planning
/gen-skeleton rag

/gen-skeleton planning example output:

from google import genai
from google.adk.agents import LlmAgent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
import asyncio

planner = LlmAgent(
    name="planner",
    model="gemini-2.5-flash",
    instruction="""You are a planning agent. Given a complex goal, decompose it
    into an ordered list of concrete subtasks. For each subtask specify:
    what to do, what input it needs, and what output it produces.""",
)

async def run(goal: str) -> str:
    session_service = InMemorySessionService()
    await session_service.create_session(
        app_name="planning-demo", user_id="user", session_id="s1"
    )
    runner = Runner(
        agent=planner, app_name="planning-demo", session_service=session_service
    )
    async for event in runner.run_async(
        user_id="user", session_id="s1",
        new_message=genai.types.Content(
            role="user", parts=[genai.types.Part(text=goal)]
        ),
    ):
        if event.is_final_response():
            return event.content.parts[0].text
    return ""

if __name__ == "__main__":
    print(asyncio.run(run("Build a customer support bot for returns, billing, and tech issues")))

Sub-agents (Preview)

@architect — recommends an optimal pattern combination for your problem.

Input:  natural-language problem description

Output:
  - Problem analysis (scope, constraints)
  - Recommended patterns with rationale
  - How the patterns combine in a system design
  - Next step: /gen-skeleton <primary-pattern>

@architect "I need to build a customer support bot that learns from feedback"

@reviewer — reviews agent code for pattern compliance and SDK conventions.

Input:  Python agent code (paste directly)

Output:
  - Pattern compliance checklist (12 core/state/reliability patterns)
  - SDK convention check (LlmAgent, Runner, InMemorySessionService, google-genai)
  - Issues found with file/line references
  - Concrete fix recommendations

@reviewer
# then paste your code

Recommended workflow:

1. @architect → get pattern recommendations
2. /gen-skeleton <pattern> → generate code skeleton
3. @reviewer → verify implementation

---

Extension Management

# Check installed extensions and their skills
gemini extensions list

# Update to latest version
gemini extensions update agentic-design-patterns

# Disable without uninstalling
gemini extensions disable agentic-design-patterns

# Uninstall
gemini extensions uninstall agentic-design-patterns

Platform Compatibility

Platform	Installation	Activation
Gemini CLI	gemini extensions install <url>	Semantic — model reads description autonomously
Antigravity	Copy skills/ to .agents/skills/	Keyword pattern matching
Claude Code	Symlink skills/ to .claude/skills/	Semantic judgment + slash commands

For Antigravity and Claude Code, use the Skills-only version.

Extension Structure

agentic-design-patterns/
├── gemini-extension.json     ← Extension manifest (v2.2.4)
├── GEMINI.md                 ← Global context: pattern guide, model guide, tech decisions
├── mcp_server.py             ← Skill-search MCP server (list_patterns, get_skill, search_skills)
├── commands/
│   ├── gen-skeleton.toml    ← /gen-skeleton <pattern> — generate code skeleton
│   └── pattern-summary.toml ← /pattern-summary [filter] — browse patterns
├── agents/
│   ├── architect.md         ← Recommend optimal pattern combinations
│   └── reviewer.md          ← Review code for pattern compliance
└── skills/                   ← 28 skill definitions
    ├── planning/
    │   └── SKILL.md
    ├── rag/
    │   └── SKILL.md
    └── ...

Trigger Coverage

Language	Count
English	474
Korean	337
Japanese	284
Chinese	281
Total	1,376

Model Reference

Task Type	Recommended Model	Thinking Budget
Simple pipelines — prompt-chaining, routing	gemini-2.5-flash-lite	Not supported
Medium complexity — tool-use, RAG, parallelization	gemini-2.5-flash	Dynamic (leave unset)
Complex reasoning — planning, reasoning, evaluation	gemini-2.5-flash or gemini-2.5-pro	Set high
Large-scale coordination — multi-agent, a2a	gemini-2.5-pro	Set high

Thinking Budget is an adjustable reasoning depth parameter available on Flash and Pro models (not Flash-Lite). Leave it unset for most tasks — the model decides dynamically.

Source

Based on "Agentic Design Patterns" by Antonio Gulli (424 pages, 21 chapters + 6 appendices).

License

MIT License — free to use, modify, and distribute.