One-Line Summary: The 2025 "second-wave" of agent SDKs — OpenAI Agents SDK, Mastra (TypeScript-first), and Google ADK (Agent Development Kit) — converged on a similar shape: opinionated agent + handoff + guardrail primitives sitting between bare API calls and a full framework like LangGraph; a useful comparison if you're picking an SDK.
Prerequisites: Harness vs. framework vs. SDK, claude agent sdk overview
What Are These SDKs?
By 2025 every major frontier-model lab and several independents had shipped their own agent SDK. The category was crowded; the differences were subtle. Three worth knowing:
OpenAI Agents SDK (released March 2025): Python and TypeScript. Primitives: Agent (an LLM with instructions and tools), Runner (the execution loop), handoffs (cross-agent dispatch), input_guardrails / output_guardrails, tracing. Approximately the OpenAI counterpart to the Claude Agent SDK.
Mastra (TypeScript-first, 2024–): Built by the team behind Gatsby. Strong TypeScript ergonomics, RAG primitives, integrations with Vercel and Next.js. Targets web developers building agent features inside Node apps.
Google ADK (Agent Development Kit, late 2024–): Python. Tight integration with Gemini and Google Cloud. Strong tools for evaluation, deployment to Cloud Run, multi-modal support. The most enterprise-shaped of the three.
How They Compare
| Dimension | OpenAI Agents SDK | Mastra | Google ADK |
|---|---|---|---|
| Languages | Python, TypeScript | TypeScript-first | Python |
| Primary models | OpenAI (GPT family) | Provider-agnostic | Google (Gemini) |
| Multi-agent | Handoffs | Yes | Yes |
| Guardrails | First-class | DIY | First-class |
| Tracing | First-class | Yes | First-class |
| Deployment story | DIY | Vercel-friendly | Cloud Run |
| Best for | OpenAI-first apps | Web/Next.js apps | Google Cloud + Gemini |
| Learning curve | Moderate | Gentle | Steeper |
When to Pick One
OpenAI Agents SDK when you're already on OpenAI and want their idiomatic agent shape. Especially if you value OpenAI's tracing UI and want guardrails as a first-class primitive.
Mastra when you're building a Node/Next.js application that needs agent capabilities and you want TypeScript ergonomics throughout. Good fit if your application boundary is more important than your agent boundary.
Google ADK when you're committed to Gemini and Google Cloud. Strongest fit for Cloud Run deployment, Gemini-multimodal use cases, and enterprise environments already on GCP.
None of them when you've outgrown SDK shape and need a full framework (LangGraph) or are extending an existing harness (Claude Code + plugins).
Why It Matters
The SDK proliferation tells you something about the market: every major model lab now believes the SDK layer is competitively important. The implications for course-takers:
- The Claude Agent SDK is not the only option for building agent applications, even if it is what most of this course's examples assume.
- SDK choice often follows model choice, but doesn't have to. You can use the OpenAI Agents SDK with non-OpenAI models (with effort), Mastra with any provider, ADK with Gemini-equivalents.
- SDK choice locks you in less than framework choice. SDKs are smaller; switching between them is mostly rewriting the agent definition layer, not the application.
Connections to Other Concepts
harness-vs-framework-vs-sdk.md— The category these occupy.claude-agent-sdk-overview.md— The Anthropic counterpart.langgraph-vs-autogen-vs-crewai.md— The framework alternative.choosing-your-harness-stack.md— Capstone.
Further Reading
- OpenAI, Agents SDK documentation.
- Mastra documentation.
- Google, Agent Development Kit documentation.