SDK
Platforms:
claudeopenaigeminim365-copilot
What an SDK Is
Section titled “What an SDK Is”An SDK (Software Development Kit) is a framework or toolkit that provides abstractions for building AI workflows in code. Where APIs give you raw access to models and services, SDKs give you patterns and structure — handling orchestration logic, tool use, memory management, and multi-agent coordination so you don’t wire everything from scratch.
Think of it this way: an API lets you call a model. An SDK lets you build an agent that calls models, uses tools, manages state, and coordinates with other agents — with the plumbing handled for you.
Key Characteristics
Section titled “Key Characteristics”- Higher-level abstractions — wrap raw API calls with patterns for tool use, memory, multi-turn conversations, and agent loops
- Handle orchestration logic — manage the planning, routing, and error recovery that make agents work reliably
- Opinionated about patterns — provide built-in support for agent loops, handoffs between agents, guardrails, and structured outputs
- Multi-language support — most SDKs offer Python as a primary language, with TypeScript/JavaScript as a secondary option
When to Use It
Section titled “When to Use It”Use SDKs when:
- You’re building agents or multi-step workflows in code and want established patterns rather than building from scratch
- You need tool use orchestration — the SDK handles the loop of “model decides to call a tool → tool executes → result goes back to model”
- You’re coordinating multiple agents that need to hand off work to each other
- You want guardrails, memory management, or structured output validation built into the framework
- You’re moving beyond a single API call into workflows that require state, iteration, and decision-making
Example
Section titled “Example”Using the Claude Agent SDK to build a research agent that takes a topic, searches the web, reads relevant pages, evaluates source quality, and produces a structured report — the SDK handles the agent loop (plan → act → observe → repeat), tool execution, and conversation state, while you define the tools and instructions.
Or using LangGraph to orchestrate a content pipeline where a research agent hands off findings to a writing agent, which hands off a draft to an editing agent — each with different tools and instructions, coordinated through a graph-based workflow.
Frameworks
Section titled “Frameworks”| Framework | Provider | Languages | Links |
|---|---|---|---|
| Claude Agent SDK | Anthropic | Python, TypeScript | Playbook guide · Docs |
| OpenAI Agents SDK | OpenAI | Python, TypeScript | Playbook guide · Docs |
| Google Agent Development Kit (ADK) | Python | Playbook guide · Docs | |
| LangGraph | LangChain | Python, JavaScript | langchain.com/langgraph |
| Microsoft Agent Framework | Microsoft | Python, .NET (C#) | GitHub |
| Microsoft 365 Agents SDK | Microsoft | .NET, Python, TypeScript | Playbook guide · GitHub · Docs |
| CrewAI | CrewAI | Python | crewai.com |
Agent-to-Agent (A2A) Protocol
Section titled “Agent-to-Agent (A2A) Protocol”The Agent-to-Agent (A2A) protocol is an open standard for agent interoperability. It defines how agents built with different frameworks can discover each other, negotiate capabilities, and collaborate on tasks — regardless of which SDK or platform they were built on.
A2A complements MCP: where MCP connects agents to tools and data, A2A connects agents to other agents.
A2A specification and documentation
Key Concepts
Section titled “Key Concepts”Agent loop — The core pattern in most SDKs: the model receives a goal, decides what to do, calls tools, observes results, and repeats until the goal is met or a stop condition is reached.
Tool definition — SDKs provide structured ways to define tools the agent can call — functions, APIs, file operations, web search — with schemas that tell the model what each tool does and what parameters it accepts.
Handoffs — In multi-agent workflows, one agent passes control (and context) to another. SDKs formalize this with handoff protocols that manage what information transfers between agents.
Guardrails — Built-in safety checks that validate inputs and outputs at each step. SDKs let you define rules (content filters, output validators, scope limiters) that run automatically during agent execution.
Memory and state — SDKs manage conversation history, tool call results, and workflow state so the agent can reference prior steps. Some support persistent memory across sessions.
Structured outputs — Many SDKs integrate schema validation so agent responses conform to expected formats — critical for pipelines where one agent’s output feeds into the next.
Platform Implementations
Section titled “Platform Implementations”| Platform | How It Works | SDK Reference |
|---|---|---|
| Claude (Agent SDK) | Python and TypeScript SDK for building agents with tool use, handoffs, and guardrails; integrates with MCP servers | Docs · Python SDK · TypeScript SDK |
| OpenAI (Agents SDK) | Python and TypeScript SDK for building agents with function calling, handoffs, and tracing; includes built-in guardrails | Docs · Python SDK · TypeScript SDK |
| Google (ADK) | Python SDK for building agents on Vertex AI; supports tool use, multi-agent orchestration, and Google Cloud integrations | Docs · Python SDK |
| Microsoft (M365 Agents SDK) | .NET, Python, and TypeScript SDK for building agents that deploy to Microsoft 365 surfaces; integrates with Copilot | Docs · GitHub |
Relationship to Other Blocks
Section titled “Relationship to Other Blocks”SDK is the code-first orchestration layer:
- Model is what SDKs orchestrate — they manage the model’s decision-making loop
- API is what SDKs call under the hood — they abstract raw API calls into higher-level patterns
- Prompts are embedded in SDK agent definitions as system instructions and tool descriptions
- Context is managed by the SDK through memory, conversation history, and tool results
- Skills are conceptually similar to SDK tool definitions — both package capabilities for the agent to invoke
- Agents are what SDKs build — the agent concept comes to life through SDK code
- MCP servers are a common tool type in SDKs — agents use MCP to connect to external systems
- CLI tools like Claude Code are built on SDKs and use them to power agent capabilities in the terminal
Related
Section titled “Related”- Agentic Building Blocks — SDK in the context of all building blocks
- API — the raw programmatic interfaces that SDKs abstract over
- Agents — the building block that SDKs implement in code
- Agent Capability Patterns — architectural patterns that SDKs help implement
- Multi-Agent Collaboration — patterns for coordinating multiple agents
- MCP — the protocol that connects SDK-built agents to tools and data
- AI Use Cases — what teams build with these blocks
- Coding Use Cases — code-first AI workflows
- Workflow Architecture Patterns — architectural patterns from augmented LLMs to autonomous agents
- CLI — terminal-native interfaces that use SDKs under the hood
- Platforms — platform-specific SDK guides