aws-samples · faridun-ag2 · Mar 27, 2026
diff --git a/agents-and-function-calling/open-source-agents/README.md b/agents-and-function-calling/open-source-agents/README.md
@@ -2,6 +2,25 @@
 
 
 
+## Contents
+
+### [AG2](ag2/)
+Multi-agent framework with native Amazon Bedrock support.
+Examples: single agent, tool use, multi-agent GroupChat — all using
+`LLMConfig(api_type="bedrock")` with no wrapper libraries.
+
+### [CrewAI](crew.ai/)
+CrewAI agent examples with Amazon Bedrock.
+
+### [LangChain](LangChain/)
+LangChain agent examples with Amazon Bedrock.
+
+### [LangGraph](langgraph/)
+LangGraph agent examples with Amazon Bedrock.
+
+### [LlamaIndex](llamaindex/)
+LlamaIndex agent examples with Amazon Bedrock.
+
 ## Contributing
 
 We welcome community contributions! Please ensure your sample aligns with AWS [best practices](https://aws.amazon.com/architecture/well-architected/), and please update the **Contents** section of this README file with a link to your sample, along with a description.
diff --git a/agents-and-function-calling/open-source-agents/ag2/README.md b/agents-and-function-calling/open-source-agents/ag2/README.md
@@ -0,0 +1,43 @@
+# AG2 with Amazon Bedrock
+
+[AG2](https://ag2.ai/) (formerly AutoGen) is an open-source multi-agent framework with
+**native Amazon Bedrock support**. Unlike other frameworks that require wrapper libraries,
+AG2 connects to Bedrock directly via `LLMConfig(api_type="bedrock")`.
+
+## Examples
+
+| Notebook | Description |
+|----------|-------------|
+| [ag2-single-agent-bedrock.ipynb](ag2-single-agent-bedrock.ipynb) | Basic single agent with Bedrock |
+| [ag2-tool-use-bedrock.ipynb](ag2-tool-use-bedrock.ipynb) | Function calling with `register_for_llm` |
+| [ag2-multi-agent-bedrock.ipynb](ag2-multi-agent-bedrock.ipynb) | Multi-agent GroupChat orchestration |
+
+## Why AG2 + Bedrock?
+
+- **Native support**: `LLMConfig(api_type="bedrock")` — no LangChain or wrapper needed
+- **AWS credential chain**: IAM roles, environment variables, or `~/.aws/credentials`
+- **All Bedrock models**: Claude, Llama, Mistral, Titan, Command R+
+- **Multi-agent**: GroupChat with automatic speaker selection
+- **500K+ monthly PyPI downloads**: Active community with frequent releases
+
+## Quick Start
+
+```bash
+pip install ag2[openai]
+```
+
+```python
+from autogen import AssistantAgent, UserProxyAgent, LLMConfig
+
+llm_config = LLMConfig(
+    api_type="bedrock",
+    model="anthropic.claude-3-sonnet-20240229-v1:0",
+    aws_region="us-east-1",
+)
+```
+
+## Resources
+
+- [AG2 Documentation](https://docs.ag2.ai/)
+- [AG2 GitHub](https://github.com/ag2ai/ag2)
+- [AG2 Bedrock Guide](https://docs.ag2.ai/docs/user-guide/models/amazon-bedrock)
diff --git a/agents-and-function-calling/open-source-agents/ag2/ag2-multi-agent-bedrock.ipynb b/agents-and-function-calling/open-source-agents/ag2/ag2-multi-agent-bedrock.ipynb
@@ -0,0 +1,243 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# AG2 Multi-Agent GroupChat with Amazon Bedrock\n",
+    "\n",
+    "## Overview\n",
+    "\n",
+    "[AG2](https://ag2.ai/) (formerly AutoGen) provides a powerful **GroupChat** feature for multi-agent orchestration. The `GroupChatManager` uses the LLM to automatically select the next speaker based on conversation context — no hardcoded routing graphs or handoff logic required.\n",
+    "\n",
+    "In this notebook, we'll create a multi-agent research team using AG2's GroupChat with Amazon Bedrock as the LLM backend.\n",
+    "\n",
+    "## Context\n",
+    "\n",
+    "AG2's GroupChat is its flagship multi-agent feature. Unlike frameworks that require explicit handoff definitions or routing graphs, AG2's `GroupChatManager` uses the LLM itself to determine which agent should speak next based on:\n",
+    "- Each agent's name and system message\n",
+    "- The current conversation history\n",
+    "- The task at hand\n",
+    "\n",
+    "This makes it easy to add or remove agents without rewriting orchestration code. Combined with native Bedrock support, you get enterprise-grade multi-agent systems with AWS IAM authentication, VPC endpoints, and CloudTrail logging — all without wrapper libraries.\n",
+    "\n",
+    "## Prerequisites\n",
+    "\n",
+    "- An AWS account with Amazon Bedrock model access enabled\n",
+    "- Python 3.10+\n",
+    "- AWS credentials configured (IAM role, environment variables, or `~/.aws/credentials`)\n",
+    "- Model access granted for `anthropic.claude-3-sonnet-20240229-v1:0` in your AWS region"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Setup\n",
+    "\n",
+    "Install the AG2 package."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%pip install -q ag2[openai]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Code\n",
+    "\n",
+    "### Configure Bedrock\n",
+    "\n",
+    "Set up the native Bedrock connection."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from autogen import AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager, LLMConfig\n",
+    "\n",
+    "# Native Bedrock support — no OpenAI key needed\n",
+    "# ---- ⚠️ Update region for your AWS setup ⚠️ ----\n",
+    "llm_config = LLMConfig(\n",
+    "    api_type=\"bedrock\",\n",
+    "    model=\"anthropic.claude-3-sonnet-20240229-v1:0\",\n",
+    "    aws_region=\"us-east-1\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Create Specialist Agents\n",
+    "\n",
+    "We'll create a team of three specialist agents, each with a distinct role:\n",
+    "- **Researcher**: Gathers information and provides factual analysis\n",
+    "- **Writer**: Creates clear, well-structured content from research findings\n",
+    "- **Critic**: Reviews content for accuracy and completeness, and terminates when satisfied"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "with llm_config:\n",
+    "    researcher = AssistantAgent(\n",
+    "        name=\"Researcher\",\n",
+    "        system_message=(\n",
+    "            \"You are a research analyst. Search for information and provide \"\n",
+    "            \"factual analysis. Focus on key data points and trends. \"\n",
+    "            \"Cite sources when possible.\"\n",
+    "        ),\n",
+    "    )\n",
+    "    writer = AssistantAgent(\n",
+    "        name=\"Writer\",\n",
+    "        system_message=(\n",
+    "            \"You are a technical writer. Take research findings and create \"\n",
+    "            \"clear, well-structured summaries for business stakeholders. \"\n",
+    "            \"Use bullet points and concise language.\"\n",
+    "        ),\n",
+    "    )\n",
+    "    critic = AssistantAgent(\n",
+    "        name=\"Critic\",\n",
+    "        system_message=(\n",
+    "            \"You review content for accuracy, completeness, and clarity. \"\n",
+    "            \"Provide constructive feedback. When the output meets quality \"\n",
+    "            \"standards, say TERMINATE to end the conversation.\"\n",
+    "        ),\n",
+    "    )\n",
+    "    user_proxy = UserProxyAgent(\n",
+    "        name=\"user\",\n",
+    "        human_input_mode=\"NEVER\",\n",
+    "        max_consecutive_auto_reply=0,\n",
+    "    )"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Set Up GroupChat\n",
+    "\n",
+    "The `GroupChat` collects agents into a group, and the `GroupChatManager` orchestrates the conversation. The `speaker_selection_method=\"auto\"` setting lets the LLM decide which agent speaks next based on context.\n",
+    "\n",
+    "The `max_round` parameter limits the total number of turns to prevent runaway conversations."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "groupchat = GroupChat(\n",
+    "    agents=[user_proxy, researcher, writer, critic],\n",
+    "    messages=[],\n",
+    "    max_round=8,\n",
+    "    speaker_selection_method=\"auto\",\n",
+    ")\n",
+    "manager = GroupChatManager(groupchat=groupchat, llm_config=llm_config)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Run the Multi-Agent Conversation\n",
+    "\n",
+    "The user proxy sends the initial request to the GroupChatManager, which then automatically routes messages to the appropriate agents. Watch the conversation to see how:\n",
+    "1. The **Researcher** gathers information\n",
+    "2. The **Writer** creates a structured summary\n",
+    "3. The **Critic** reviews and provides feedback or approves"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "result = user_proxy.initiate_chat(\n",
+    "    manager,\n",
+    "    message=(\n",
+    "        \"Research the current state of generative AI adoption in enterprise. \"\n",
+    "        \"Write a brief executive summary with key trends and challenges.\"\n",
+    "    ),\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Understanding the Output\n",
+    "\n",
+    "The GroupChatManager automatically routed the conversation through the agents:\n",
+    "- The **Researcher** provided data points and analysis\n",
+    "- The **Writer** structured the findings into an executive summary\n",
+    "- The **Critic** reviewed the output and said TERMINATE when satisfied\n",
+    "\n",
+    "The key advantage of AG2's GroupChat: the `GroupChatManager` uses the LLM to automatically select the next speaker based on conversation context. You don't need to define explicit routing logic or handoff patterns."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Best Practices\n",
+    "\n",
+    "- **Speaker selection**: Use `\"auto\"` for LLM-based routing, `\"round_robin\"` for predictable sequential flow\n",
+    "- **Max rounds**: Set `max_round` to prevent runaway conversations — 6-10 is a good starting range\n",
+    "- **Termination**: Include `TERMINATE` in one agent's system message to end gracefully\n",
+    "- **Agent count**: 3-5 agents works well; more agents increases speaker selection complexity\n",
+    "- **Distinct roles**: Give each agent a clear, non-overlapping system message for better routing\n",
+    "- **Native Bedrock advantages**: No OpenAI key, AWS IAM auth, supports Claude/Llama/Mistral/Titan, enterprise-grade security with VPC endpoints and CloudTrail logging\n",
+    "\n",
+    "## Next Steps\n",
+    "\n",
+    "- **Single agent**: See [ag2-single-agent-bedrock.ipynb](ag2-single-agent-bedrock.ipynb) for the basic setup\n",
+    "- **Tool use**: See [ag2-tool-use-bedrock.ipynb](ag2-tool-use-bedrock.ipynb) for function calling with Bedrock\n",
+    "- **AG2 GroupChat Guide**: [docs.ag2.ai/docs/user-guide/basic-concepts/orchestration/group-chat](https://docs.ag2.ai/docs/user-guide/basic-concepts/orchestration/group-chat)\n",
+    "- **AG2 Documentation**: [docs.ag2.ai](https://docs.ag2.ai/)\n",
+    "- **AG2 Bedrock Guide**: [docs.ag2.ai/docs/user-guide/models/amazon-bedrock](https://docs.ag2.ai/docs/user-guide/models/amazon-bedrock)\n",
+    "\n",
+    "## Cleanup\n",
+    "\n",
+    "No resources to clean up — this notebook uses only local compute and Bedrock API calls. To stop incurring Bedrock charges, simply stop running the notebook."
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.11.7"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}