agent-frameworks git:20261107.e294cb4A

Immutable. This exact content never changes and is served at /api/v1/blob/eb705b514f1f34fc.

---
name: agent-frameworks
description: "Build Python agents and their system prompts with Google ADK, LangGraph, LangChain, and agents-cli."
license: MIT
metadata:
  kind: collection
  author: Médéric HURIER (Fmind)
  source: github.com/fmind/dot/tree/main/skills/agent-frameworks
  created: "2026-09-16"
  updated: "2026-11-07"
---

# Agent Frameworks

Implement the existing or explicitly selected agent framework. Choosing a framework does not authorize provider calls, tracing exports, deployment, or installing every vendor skill.

## Workflow

1. **Keep the project's framework**: an existing agent stays on its framework unless the user asks to migrate it.
1. **Default new agents to Google ADK**: scaffold with agents-cli, then write SDK code per google-adk.
1. **Use LangGraph second**: choose it when ADK does not fit, for example a required non-Google stack, durable graph state, or interrupts ADK cannot express. Use LangChain only as LangGraph building blocks: start from its `create_agent`, which returns a compiled LangGraph, and hand-write graph topology only where the standard loop falls short. No other framework is a default.
1. **Read only the matching guide** and its required resources.

## Task guides

<!-- guides:start -->

- [agents-cli](references/agents-cli/GUIDE.md): Google agent scaffolding, lifecycle, and official skill discovery.
- [google-adk](references/google-adk.md): Google ADK agents, runners, tools, sessions, and evaluation.
- [langchain](references/langchain.md): LangChain create_agent, middleware, MCP tools, and structured output as LangGraph building blocks.
- [langgraph](references/langgraph.md): LangGraph state, persistence, streaming, interrupts, and durable execution.
- [system-prompts](references/system-prompts.md): Application system prompt layering, rule wording, and validated structured output.

<!-- guides:end -->