Skip to content

frameworks.md

Latest commit

 

History

History
234 lines (162 loc) · 8.51 KB

frameworks.md

File metadata and controls

234 lines (162 loc) · 8.51 KB
Raw

Agent Frameworks

AgentCore CLI supports multiple agent frameworks for template-based agent creation, plus a BYO (Bring Your Own) option for existing code.

Supported Languages

Language Supported Frameworks Runtime Notes
Python All frameworks Python 3.12+ Default language. Uses uv for dependency management.
TypeScript Strands, Vercel AI Node 22 Uses npm + tsx for the dev loop. Other frameworks are not yet available in TS.

Pass --language TypeScript to agentcore create or agentcore add agent to scaffold a TypeScript project. The framework is restricted to Strands or VercelAI; other values are rejected. See Local Development for the TS dev loop.

Available Frameworks

Framework Supported Model Providers
Strands Agents Bedrock, Anthropic, OpenAI, Gemini
LangChain_LangGraph Bedrock, Anthropic, OpenAI, Gemini
GoogleADK Gemini only
OpenAIAgents OpenAI only
VercelAI Bedrock, Anthropic, OpenAI, Gemini
BedrockManagedAgents Bedrock only (the model runs in Bedrock Managed Agents)

Runtime Input Validation

Validate agent invocation payloads before passing them to a framework. Template entrypoints validate plain prompts as strings; preserve that validation when extending a generated application, and pass only prompt text to the agent.

Framework Selection Guide

Strands Agents

AWS's native agent framework designed for Amazon Bedrock.

Best for:

  • Projects primarily using Amazon Bedrock models
  • Integration with AWS services
  • Production deployments on AWS infrastructure

Model providers: Bedrock, Anthropic, OpenAI, Gemini

Languages: Python, TypeScript

agentcore create --framework Strands --model-provider Bedrock

# TypeScript variant
agentcore create --framework Strands --model-provider Bedrock --language TypeScript

LangChain / LangGraph

Popular open-source framework with extensive ecosystem.

Best for:

  • Complex multi-step agent workflows
  • Projects requiring LangChain's extensive tool ecosystem
  • Teams already familiar with LangChain

Model providers: Bedrock, Anthropic, OpenAI, Gemini

agentcore create --framework LangChain_LangGraph --model-provider Anthropic

GoogleADK

Google's Agent Development Kit.

Best for:

  • Projects using Google's Gemini models
  • Integration with Google Cloud services

Model providers: Gemini only

agentcore create --framework GoogleADK --model-provider Gemini

OpenAIAgents

OpenAI's native agent framework.

Best for:

  • Projects using OpenAI models exclusively
  • Simple agent workflows with OpenAI's function calling

Model providers: OpenAI only

agentcore create --framework OpenAIAgents --model-provider OpenAI --api-key sk-...

Vercel AI SDK

Vercel's AI SDK for building AI-powered applications.

Best for:

  • Full-stack AI applications with streaming support
  • Projects using Vercel's ecosystem
  • TypeScript-first agent development

Model providers: Bedrock, Anthropic, OpenAI, Gemini

Languages: Python, TypeScript

agentcore create --framework VercelAI --model-provider Bedrock

# TypeScript variant
agentcore create --framework VercelAI --model-provider Bedrock --language TypeScript

Bedrock Managed Agents

A Runtime environment for Amazon Bedrock Managed Agents (BMA). BMA runs the agent loop (Codex) in the service and sends the session's commands to this Runtime. The Runtime runs no model code.

Best for:

  • Running BMA sessions in your account, with your execution role and VPC

Model providers: Bedrock only

Languages: Python (Container build only, no memory)

BMA is a short name for the framework. The template sets an idle timeout of 1800 seconds, a maximum lifetime of 28800 seconds, and a policy that lets the Runtime connect to BMA. It adds no session storage, so the same project deploys to a microVM Runtime and to a capacity provider.

agentcore create --name MyManagedAgent --framework BedrockManagedAgents

Import from Bedrock Agents

If you have an existing Bedrock Agent, you can import its configuration and translate it into runnable Strands or LangChain/LangGraph code. The imported agent preserves your Bedrock Agent's action groups, knowledge bases, multi-agent collaboration, guardrails, prompts, and memory configuration.

# Interactive (select "Import from Bedrock Agents" in the wizard)
agentcore add agent

# Non-interactive
agentcore add agent \
  --name MyAgent \
  --type import \
  --agent-id AGENT123 \
  --agent-alias-id ALIAS456 \
  --region us-east-1 \
  --framework Strands \
  --memory none

What gets imported

The import process fetches your Bedrock Agent's full configuration and translates it into framework-specific Python code that runs on AgentCore:

  • Action groups (function-schema and built-in) become @tool decorated functions
  • Knowledge bases become retrieval tool integrations
  • Multi-agent collaboration produces separate collaborator files with recursive translation
  • Code interpreter wires to AgentCore's code_interpreter_client
  • Guardrails are configured in the model initialization
  • Prompt overrides are preserved as template variables
  • Memory integrates with AgentCore's memory service when enabled

Import options

Flag Description
--type import Use import mode (required)
--agent-id <id> Bedrock Agent ID
--agent-alias-id <id> Bedrock Agent Alias ID
--region <region> AWS region where the Bedrock Agent exists
--framework <fw> Strands or LangChain_LangGraph
--memory <opt> none, shortTerm, longAndShortTerm

Bring Your Own (BYO) Agent

For existing agent code or frameworks not listed above, use the BYO option:

agentcore add agent \
  --name MyAgent \
  --type byo \
  --code-location ./my-agent \
  --entrypoint main.py \
  --language Python

BYO Requirements

  1. Entrypoint: Your code must expose an HTTP endpoint that accepts agent invocation requests
  2. Code location: Directory containing your agent code
  3. Language: Python

BYO Options

Flag Description
--type byo Use BYO mode (required)
--code-location <path> Directory containing your agent code
--entrypoint <file> Entry file (e.g., main.py or index.ts)
--language <lang> Python

Framework Comparison

Feature Strands LangChain GoogleADK OpenAIAgents VercelAI
Multi-provider support Yes Yes No No Yes
AWS Bedrock native Yes No No No No
Tool ecosystem Growing Extensive Moderate Moderate Moderate
Memory integration Native Via libs Via libs Via libs Via libs

Protocol Compatibility

Not all frameworks support all protocol modes. MCP protocol is a standalone tool server with no framework.

Protocol Supported Frameworks
HTTP Strands, LangChain_LangGraph, GoogleADK, OpenAIAgents, VercelAI, BedrockManagedAgents
MCP None (standalone tool server)
A2A Strands, GoogleADK, LangChain_LangGraph