prepared by Claude.ai with some guidance
AI Agent SDKs & Frameworks Comparison (January 2025)
A comprehensive guide to the major cloud provider and open-source options for building AI agents.
Executive Summary
The AI agent landscape has converged rapidly in late 2024 and early 2025. All three major cloud providers now offer both open-source SDKs for building agents and managed platforms for deploying them at scale. Key trends include:
- MCP (Model Context Protocol) has become the standard for tool integration across all platforms
- A2A (Agent-to-Agent) protocol is emerging for multi-agent communication
- Model-driven approaches are replacing complex workflow orchestration
- Memory, governance, and observability are now table-stakes features
Quick Comparison
| Provider | Open Source SDK | Managed Platform | Languages | MCP Support |
|---|---|---|---|---|
| AWS | Strands Agents | Bedrock AgentCore | Python, TypeScript | ✅ Native |
| Agent Development Kit (ADK) | Vertex AI Agent Engine | Python, Java, Go, TypeScript | ✅ Native + Managed Servers | |
| Microsoft | Agent Framework | Azure AI Agent Service | .NET, Python | ✅ Native |
AWS
Strands Agents (Open Source SDK)
Strands is AWS's open-source, model-driven framework that lets the LLM decide the execution flow rather than requiring developers to define explicit workflows.
Philosophy: "Give the agent tools and a prompt, let the model figure out the rest."
Key Features:
- Minimal code required (under 10 lines to get started)
- Model-agnostic: Bedrock, Anthropic, OpenAI, Ollama, Gemini, Llama, and more
- Native MCP support for tool integration
- Multi-agent orchestration (v1.0): agents-as-tools, swarms, A2A protocol
- Edge device support for IoT/automotive use cases
- Powers Amazon Q Developer, AWS Glue, and Kiro internally
Languages: Python, TypeScript (new as of re:Invent 2025)
Installation:
pip install strands-agents strands-agents-toolsBasic Example:
from strands import Agent
from strands_tools import calculator
agent = Agent(tools=[calculator])
agent("What is the square root of 1764")Links:
- GitHub: https://github.com/strands-agents/sdk-python
- Documentation: https://strandsagents.com
- AWS Blog: https://aws.amazon.com/blogs/opensource/introducing-strands-agents-an-open-source-ai-agents-sdk/
Amazon Bedrock AgentCore (Managed Platform)
AgentCore is AWS's managed platform for deploying, operating, and governing AI agents at scale. Think of it as an "operating system for agents."
Key Services:
| Service | Purpose |
|---|---|
| AgentCore Runtime | 8-hour execution windows, session isolation, A2A protocol |
| AgentCore Gateway | Secure tool/API connections with real-time interception |
| AgentCore Policy | Natural language → Cedar policy enforcement (deterministic, outside LLM loop) |
| AgentCore Evaluations | 13 built-in evaluators for quality monitoring |
| AgentCore Memory | Episodic memory — agents learn from experiences |
| AgentCore Identity | IAM integration, custom claims for multi-tenant auth |
| AgentCore Observability | CloudWatch-integrated monitoring and tracing |
New at re:Invent 2025:
- Policy controls that enforce boundaries deterministically (not probabilistically)
- Episodic memory for learning from past interactions
- Bidirectional streaming for voice agents
- VPC, PrivateLink, and CloudFormation support
Links:
- Product Page: https://aws.amazon.com/bedrock/agentcore/
- Documentation: https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html
- Announcement: https://aws.amazon.com/blogs/aws/amazon-bedrock-agentcore-adds-quality-evaluations-and-policy-controls-for-deploying-trusted-ai-agents/
AWS Frontier Agents
A new class of autonomous agents announced at re:Invent 2025 that can work for days without human intervention:
| Agent | Purpose | Status |
|---|---|---|
| Kiro Autonomous Agent | Multi-repo code changes, learns from PR feedback, persistent context | Preview |
| Security Agent | Continuous vulnerability testing throughout SDLC | Preview |
| DevOps Agent | Incident response, root cause analysis across CloudWatch/GitHub/ServiceNow | Preview |
Google Cloud
Agent Development Kit (ADK) (Open Source SDK)
ADK is Google's open-source framework for building AI agents, now available in four languages. It powers agents in Google products like Agentspace and Customer Engagement Suite.
Philosophy: Code-first approach with fine-grained control over agent behavior.
Key Features:
- Multi-agent by design: hierarchical composition and delegation
- Rich model ecosystem: Gemini, Vertex AI Model Garden, LiteLLM for third-party models
- Native MCP support plus pre-built tools (Search, Code Exec)
- Bidirectional audio/video streaming for conversational agents
- Single-command deployment:
adk deployto Agent Engine - Plugin framework for custom logic (policy enforcement, usage tracking, self-healing)
Languages: Python (7M+ downloads), Java, Go, TypeScript
Installation:
pip install google-adkLinks:
- GitHub: https://github.com/google/adk-python
- Documentation: https://cloud.google.com/vertex-ai/docs/agent-builder
- TypeScript Announcement: https://developers.googleblog.com/introducing-agent-development-kit-for-typescript-build-ai-agents-with-the-power-of-a-code-first-approach/
- Go Announcement: https://developers.googleblog.com/en/announcing-the-agent-development-kit-for-go-build-powerful-ai-agents-with-your-favorite-languages/
Vertex AI Agent Engine (Managed Platform)
The managed runtime for deploying and scaling ADK agents in production.
Key Features:
| Feature | Description |
|---|---|
| Sessions & Memory Bank | Now GA — persistent state and user memory |
| Code Execution | Sandboxed environment for running agent-generated code |
| Agent Identity | IAM-based access control and authentication |
| A2A Protocol Support | Agent-to-agent communication |
| Observability | Traces, metrics, and playground in Cloud Console |
| Express Mode | Free tier for getting started |
Links:
- Product Page: https://cloud.google.com/vertex-ai/docs/agent-builder
- Release Notes: https://docs.cloud.google.com/agent-builder/release-notes
Google Managed MCP Servers (New!)
Google is launching fully managed, remote MCP servers for their services — no setup required, just paste a URL.
Available Servers (Preview):
- Google Maps
- BigQuery
- Compute Engine
- Kubernetes Engine
Coming Soon: Storage, databases, logging, monitoring, security services
Security: Protected by Google Cloud IAM and Model Armor (firewall for agentic workloads)
Links:
- Announcement: https://techcrunch.com/2025/12/10/google-is-going-all-in-on-mcp-servers-agent-ready-by-design/
Agent Payments Protocol (AP2)
A new open protocol for agent-led payments, designed as an extension of A2A and MCP.
Links:
- Announcement: https://cloud.google.com/blog/products/ai-machine-learning/what-google-cloud-announced-in-ai-this-month
Microsoft
Microsoft Agent Framework (Open Source SDK)
The unified successor to both Semantic Kernel and AutoGen, built by the same teams. Combines AutoGen's simple abstractions with Semantic Kernel's enterprise features.
Philosophy: Agents for dynamic/unstructured tasks, Workflows for structured multi-step processes.
Two Core Capabilities:
- AI Agents — Individual agents using LLMs, tools, and MCP servers
- Workflows — Graph-based multi-agent orchestration with checkpointing, routing, and human-in-the-loop
Key Features:
- Thread-based state management for multi-turn conversations
- Context providers for agent memory
- Middleware for intercepting agent actions
- Native MCP client support
- Checkpointing for long-running workflows
- Multi-agent orchestration patterns: sequential, concurrent, hand-off, Magentic
- Model providers: Azure OpenAI, OpenAI, Azure AI
Languages: .NET, Python
Installation:
# Python
pip install agent-framework --pre
# .NET
dotnet add package Microsoft.Agents.AIWhen to Use Agents vs Workflows:
- Agents: Autonomous decision-making, unstructured input, exploration, conversation
- Workflows: Structured processes, predefined rules, many tools (20+), complex coordination
"If you can write a function to handle the task, do that instead of using an AI agent."
Links:
- GitHub: https://github.com/microsoft/agent-framework
- Documentation: https://learn.microsoft.com/en-us/agent-framework/overview/agent-framework-overview
- Migration from Semantic Kernel: https://learn.microsoft.com/en-us/agent-framework/migration-guide/from-semantic-kernel/
- Migration from AutoGen: https://learn.microsoft.com/en-us/agent-framework/migration-guide/from-autogen/
Azure AI Agent Service (Managed Platform)
Built on OpenAI's Assistants API with Azure-native integrations.
Key Integrations:
- Cosmos DB
- Azure Functions
- Azure AI Search
- Azure Blob Storage
Links:
- Documentation: https://learn.microsoft.com/en-us/azure/ai-services/agents/
Popular Open Source Frameworks
LangGraph (LangChain)
Graph-based orchestration for stateful, multi-actor agent workflows. Very flexible for complex control flows.
Best For: Complex workflows requiring explicit control over execution paths
Links:
- GitHub: https://github.com/langchain-ai/langgraph
- Documentation: https://langchain-ai.github.io/langgraph/
CrewAI
Role-based multi-agent framework with intuitive abstractions for agent collaboration.
Best For: Team-based agent scenarios with clear role definitions
Links:
- GitHub: https://github.com/crewAIInc/crewAI
- Documentation: https://docs.crewai.com/
OpenAI Agents SDK
Lightweight Python SDK with handoffs, guardrails, and tracing built-in.
Best For: OpenAI-centric applications needing simple agent patterns
Links:
- GitHub: https://github.com/openai/openai-agents-python
- Documentation: https://platform.openai.com/docs/agents
Pydantic AI
Type-safe agent framework leveraging Pydantic for structured outputs.
Best For: Python developers wanting type safety and validation
Links:
- GitHub: https://github.com/pydantic/pydantic-ai
- Documentation: https://ai.pydantic.dev/
Vercel AI SDK
TypeScript toolkit for AI-powered applications with React, Next.js, Vue, Svelte, Node.js.
Best For: Web developers building AI features in JavaScript/TypeScript apps
Links:
- GitHub: https://github.com/vercel/ai
- Documentation: https://sdk.vercel.ai/docs
Other Notable Frameworks
| Framework | Focus | Link |
|---|---|---|
| Haystack (deepset) | RAG-heavy agent applications | https://haystack.deepset.ai/ |
| DSPy (Stanford) | Programmatic LLM pipeline optimization | https://github.com/stanfordnlp/dspy |
| AutoGPT | Autonomous self-prompting agents | https://github.com/Significant-Gravitas/AutoGPT |
| LlamaIndex | Knowledge agents connected to your data | https://www.llamaindex.ai/ |
Feature Comparison Matrix
| Feature | AWS Strands | Google ADK | MS Agent Framework | LangGraph | CrewAI |
|---|---|---|---|---|---|
| MCP Support | ✅ Native | ✅ Native | ✅ Native | ✅ Via tools | ❌ |
| A2A Protocol | ✅ | ✅ | ❌ | ❌ | ❌ |
| Multi-Agent | ✅ | ✅ | ✅ | ✅ | ✅ |
| Memory/State | ✅ | ✅ | ✅ | ✅ | ✅ |
| Streaming | ✅ | ✅ Bidirectional | ✅ | ✅ | ✅ |
| TypeScript | ✅ | ✅ | ❌ | ✅ | ❌ |
| Python | ✅ | ✅ | ✅ | ✅ | ✅ |
| .NET | ❌ | ❌ | ✅ | ❌ | ❌ |
| Go | ❌ | ✅ | ❌ | ❌ | ❌ |
| Java | ❌ | ✅ | ❌ | ❌ | ❌ |
| Model Agnostic | ✅ | ✅ | ✅ | ✅ | ✅ |
| Edge/IoT | ✅ | ❌ | ❌ | ❌ | ❌ |
| Managed Platform | Bedrock AgentCore | Vertex AI Agent Engine | Azure AI Agent Service | LangSmith | ❌ |
Decision Guide
Choose AWS Strands + AgentCore if:
- You want model-driven simplicity (let the LLM orchestrate)
- You're already on AWS infrastructure
- You need edge device deployment
- You prefer minimal boilerplate code
Choose Google ADK + Agent Engine if:
- You want the most language options (Python, Java, Go, TypeScript)
- You need managed MCP servers for Google services
- You're building conversational agents with audio/video streaming
- You want single-command deployment (
adk deploy)
Choose Microsoft Agent Framework if:
- You're a .NET shop or need C# support
- You're migrating from Semantic Kernel or AutoGen
- You need explicit workflow control alongside agents
- You want enterprise features like thread-based state management
Choose LangGraph if:
- You need maximum flexibility in control flow
- You're already using LangChain
- You want explicit graph-based orchestration
Choose CrewAI if:
- You're building role-based multi-agent teams
- You want intuitive, high-level abstractions
- You prefer simplicity over configurability
Key Protocols & Standards
MCP (Model Context Protocol)
Developed by Anthropic, now donated to Linux Foundation. The emerging standard for connecting AI agents to tools and data sources.
Links:
- Specification: https://modelcontextprotocol.io/
- GitHub: https://github.com/modelcontextprotocol
A2A (Agent-to-Agent Protocol)
Google-led protocol for agent-to-agent communication. Supported by AWS and Google platforms.
Links:
- Specification: https://github.com/google/A2A
Summary
The AI agent ecosystem has matured significantly. All major cloud providers now offer comparable capabilities:
- Open-source SDKs for local development and flexibility
- Managed platforms for production deployment with governance
- Native MCP support for standardized tool integration
- Multi-agent orchestration for complex workflows
- Memory and state management for personalized experiences
- Policy and governance controls for enterprise compliance
The choice between platforms often comes down to existing cloud infrastructure, language preferences, and specific feature needs like edge deployment (AWS) or managed MCP servers (Google).
