Google ADK (Python) vs LangGraph
Both are graph runtimes for agents. Google ADK 2.0 adds a Workflow graph next to coordinator agents with subagents, and ships a CLI, dev web UI and eval runner; it is optimized for Gemini and Google Cloud deployment. LangGraph is a lower-level state graph with checkpoints, interrupts and time travel, deployed through LangSmith or self-hosted.
Facts side by side
| Fact | Google ADK (Python) | LangGraph |
|---|---|---|
| Type | Framework | Framework |
| Languages / SDKs | Python | Python |
| License | Apache-2.0 | MIT |
| Pricing model | Open core | Open core |
| Orchestration pattern | Graph | Graph |
| GitHub stars | 21,686 (as of 2026-09-30) | 42,511 (as of 2026-09-30) |
| GitHub forks | 4,084 | 7,201 |
| Last push | 2026-09-30 | 2026-09-29 |
| Latest release | v2.10.0 | cli==0.4.32.dev0 |
| Repository | google/adk-python | langchain-ai/langgraph |
| Website | adk.dev | www.langchain.com |
| Documentation | adk.dev | docs.langchain.com |
| Last verified | 2026-09-30 | 2026-09-30 |
| MCP support | Yes (checked 2026-09-30) | Yes (checked 2026-09-30) |
| A2A support | Partial (checked 2026-09-30) | Yes (checked 2026-09-30) |
| AG-UI support | Partial (checked 2026-09-30) | Partial (checked 2026-09-30) |
| Install verified | 2026-09-30 | 2026-09-30 |
Choose Google ADK (Python) if
- You deploy on Google Cloud and want the documented path to the managed Agent Runtime or Cloud Run.
- You want built-in tooling in the package:
adk run,adk webandadk eval. - You need sibling SDKs in Java or Go with the same concepts (see the Google ADK Java and Go tool pages).
Choose LangGraph if
- You need stable human-in-the-loop and A2A today: ADK labels tool confirmation and Python A2A as experimental.
- You want replay and fork from any checkpoint (time travel) and per-key state reducers.
- You prefer no tie to one cloud for managed hosting; LangGraph can run standalone or on LangSmith.
Migration notes
Concept mapping is close. An ADK Workflow node corresponds to a LangGraph node; ADK edges with routing, fan-out and loops correspond to conditional edges and parallel branches; ADK Session state corresponds to graph state saved by a checkpointer, and ADK's cross-session memory service to a LangGraph store. Coordinator agents with subagents (chat, task, single-turn modes) have no built-in equivalent in LangGraph and become subgraphs or tool-wrapped agents. Human review moves between ADK tool confirmation (experimental; not supported with database or Vertex AI session services) and LangGraph interrupt(). Teams coming from ADK 1.x should read the ADK 2.0 notes first: agents became graph nodes and some legacy overrides are ignored.
FAQ
Is Google ADK tied to Gemini?
No. The README describes it as model-agnostic though optimized for Gemini.
Which has better protocol coverage?
Both document MCP. ADK can consume MCP servers and publish an agent as one. A2A in ADK Python is labelled experimental (recorded as partial); LangGraph's A2A and MCP endpoints come from Agent Server. AG-UI is partial for both, via adapters maintained in the AG-UI repository.
Are both free to use?
ADK is Apache-2.0 and LangGraph is MIT. Managed hosting is paid on both sides: Google Cloud's Agent Runtime beyond its no-cost tier, and LangSmith Deployment from the Plus plan.