Our assessment
LangChain is where most teams start, and the reason is the integration layer: one interface over every major model provider, vector store and tool, so swapping Anthropic for OpenAI is a small code change rather than a rewrite. Version 1 narrowed the surface to create_agent (a minimal agent harness you extend with middleware for guardrails, retries and routing) which answers the long-standing complaint that it had too many abstractions.
Under the hood its agents now run on LangGraph, so you inherit durable execution and human-in-the-loop when you need them without starting at the lower level. The trade-off is that you are buying into an ecosystem: the debugging and evaluation story is LangSmith, which is a paid product, and the abstractions still add a layer between you and the model API that some teams eventually remove.
Pick it if you want to move fast across providers and lean on a huge community of examples. If your app is a single model call with a few tools, the provider SDK is simpler. If you need explicit control over a long-running workflow, go straight to LangGraph.
- Best for
- Prototyping and shipping LLM apps across many providers with a large ecosystem of integrations.
- Watch out for
- Abstraction overhead for simple apps; serious observability means adopting LangSmith.
- Consider instead
- LangGraph for explicit, stateful workflows · Pydantic AI for a typed, lighter-weight framework.
How it compares in Agent frameworks
History
Tracked since Oct 13, 2026.
| Snapshot | Rank | Adoption | Entry price |
|---|---|---|---|
| Nov 20, 2026 | #2 same | 96 | Free |
| Nov 12, 2026 | #2 same | 96 | Free |
| Oct 29, 2026 | #2 same | 96 | Free |
| Oct 15, 2026 | #2 same | 96 | Free |
| Oct 13, 2026 | #2 | 96 | Free |
Key facts
| Category | Agent frameworks |
| Maker | LangChain |
| License | MIT |
| Self-host | Yes, it's a library |
| Languages | Python (LangChain.js for TypeScript) |
| Repository created | Oct 2022 |
| GitHub stars | 147.8k (Oct 15, 2026) |
| Latest release | langchain-core 1.6.10 (Oct 12, 2026) |
| Agent API | create_agent + middleware (v1) |
| Runtime | Built on LangGraph |
| Model-agnostic | Yes, OpenAI, Anthropic, Google and more |
| Commercial layer | LangSmith (tracing, evals, deployment) |
