Official Agent SDKs
Third-party frameworks like LangGraph and CrewAI (see Agent Frameworks Compared) sit on top of whichever model you call. In 2026, both Anthropic and OpenAI started shipping their own official, first-party agent SDKs instead - production infrastructure with harness-level concerns (memory, tools, permissions, orchestration) built in, maintained by the same team that ships the model.
Claude Agent SDK
The Claude Agent SDK exposes the same underlying harness that Claude Code itself runs on, as a library you can build your own agents on top of. Rather than assembling tool dispatch, permission checks, and context management from scratch, you get Anthropic's own production-tested implementation of those layers.
What it gives you
- • Context and memory management out of the box
- • A permission system for what tools an agent may call and when
- • Subagent orchestration for splitting work across specialised agents
- • The same battle-tested loop that powers Claude Code
Trade-offs
- • Tied to Claude models
- • Less flexible than a graph-based framework for highly custom control flow
- • Newer and smaller community than LangChain's ecosystem
OpenAI AgentKit
OpenAI AgentKit bundles the pieces most teams previously hand-rolled around the Assistants and Responses APIs - tool definitions, state, and orchestration - into a single first-party toolkit, positioned as the default way to build production agents on OpenAI models without adopting a third-party framework.
Official SDK vs Third-Party Framework
| Official Agent SDK | Third-party framework | |
|---|---|---|
| Model support | One provider's models only | Usually model-agnostic |
| Control flow | Opinionated, less to configure | Highly customisable (graphs, actors, events) |
| Maintenance | Matches the model provider's own release cycle | Independent open-source project |
| Best for | Teams committed to one provider, want the reference implementation | Multi-model routing, complex branching workflows |
When to Use Which
If you're building on a single provider and want the harness that team itself uses in production, start with the official SDK - it is usually the fastest path to something reliable. Reach for a third-party framework like LangGraph or CrewAI when you need model-agnostic routing, a control-flow shape the official SDK doesn't support well, or you're standardising tooling across teams that use different providers.
Checklist: Do You Understand This?
- Can you explain what distinguishes an official Agent SDK from a third-party framework like LangGraph?
- Do you know which harness layers (see Harness Engineering) an official SDK typically provides out of the box?
- Can you state one concrete reason to pick a third-party framework over an official SDK?