LangGraph
The best open-source runtime for agents that must survive a crash, and it costs nothing. Read the 6.4 as breadth, not quality: LangGraph does one layer extremely well and ships almost nothing above or below it.
PRICING
| LangGraph (open source) | $0 | MIT licence, current release 1.2.11. Python and JavaScript. No feature gating, no telemetry requirement, no usage cap. LangGraph 1.0 is designated LTS and carries a no-breaking-changes commitment until 2.0 |
| LangSmith Developer (optional) | $0 | 1 seat, 5,000 traces per month, 14-day retention. Only needed if you want the graph-aware trace UI |
| LangSmith Plus (optional) | $39/seat/mo | Unlocks LangSmith Deployment, the managed runtime formerly called LangGraph Platform. Traces meter on top; deployment compute meters in LangChain Compute Units |
The framework and the platform are separate purchases and the naming has moved. LangGraph Platform was renamed LangSmith Deployment in October 2025, and the CLI config file was renamed langgraph.json to langsmith.json to match. Nothing in the open-source package requires an account: you can run LangGraph in production having never signed up for anything. Organisations still on the previous deployment pricing keep it until 01/10/2026.
checked 2026-09-04 · vendor pricing page
Element scores
Strengths
Durable execution is the reason to pick this over writing your own loop. A checkpointer persists the full graph state after every node, so a run that dies halfway resumes from the last completed step instead of re-running — and with frontier tokens costing what they cost, that is money, not neatness. The state model is explicit: you declare a typed schema with reducers, and you can inspect, log and replay any point in a run's history. Human-in-the-loop is built into the runtime rather than bolted on, with interrupt() suspending a graph for as long as the approval takes. Multi-agent shapes are all covered — supervisor, swarm, hierarchical, arbitrary networks — with prebuilt packages for the common two. The licence is MIT with no feature gating, the 1.0 release in October 2025 is designated LTS with no breaking changes promised until 2.0, and the repo is busy: 41,054 stars and a push the day before this audit. Klarna, Uber, LinkedIn, Replit and Elastic are named production users on the vendor's own pages.
Honest dings
It does very little for you by default and it fails quietly when you assume otherwise. Three independent 2026 production write-ups land on the same failures. Checkpoints are a schema contract with no migration tooling — add a field to your state and in-flight threads stop deserialising, which one team fixed by hand-patching the checkpoint database on a Friday night. The default reducer is overwrite, so two nodes writing the same key silently drops one, and TypedDict validates nothing at runtime. The SQLite checkpointer locks under concurrency and belongs on your laptop only. Checkpoint writes are synchronous and in the hot path, one team reports a 14-node graph producing 180KB blobs and a table hitting 10GB in three weeks, and there is no built-in TTL. The checkpoint and your node's side effect are not transactionally linked, so resume means duplicate API calls unless every node carries its own idempotency key. Above the orchestration layer the box is close to empty: no evals, no retrieval, no prompt management, no guardrails beyond approval gates — each of those is a separate purchase from the same vendor, and LangSmith is where the graph-aware tracing actually lives. MCP is real but sits in the langchain package as a beta namespace, and there is no A2A. TypeScript lags Python on features.
Sources (16) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- docs.langchain.com/oss/python/langgraph/overview — Official scope statement: low-level orchestration runtime, explicitly does not abstract prompts or architecture; how LangChain, LangGraph, Deep Agents, LangSmith Observability and LangSmith Deployment divide the stack; named users Klarna, Uber, J.P. Morgan (accessed 2026-09-04)
- github.com/langchain-ai/langgraph — MIT licence, 41,054 stars / 6,933 forks / 749 open issues / 179 watchers, last push 03/09/2026, repo created 09/08/2023; README scope and named production users Klarna, Replit, Elastic (accessed 2026-09-04)
- pypi.org/pypi/langgraph/json — Current release 1.2.11, licence MIT, Development Status 5 - Production/Stable, Python 3.10-3.13, depends on langchain-core >=1.4.7 and langgraph-checkpoint >=4.1.0 (accessed 2026-09-04)
- docs.langchain.com/oss/python/langgraph/durable-exe… — Official persistence model: checkpointers for thread-scoped state, stores for cross-thread memory; documented failure modes incl. InMemorySaver losing everything on restart, unbounded checkpoint growth with no built-in retention, subgraph checkpoint namespacing (accessed 2026-09-04)
- docs.langchain.com/oss/python/langgraph/stores — Store semantic search: index config takes an embed function, dims and fields; store.search accepts a natural-language query; production backends PostgresStore, MongoDBStore, RedisStore, UpstashStore (accessed 2026-09-04)
- docs.langchain.com/oss/python/langchain/mcp — MCP support lives in the langchain[mcp] extra via MCPAdapter on FastMCP, requires langchain>=1.4.0, namespace is beta and raises LangChainBetaWarning; stdio, streamable HTTP and in-process transports; OAuth 2.1 (accessed 2026-09-04)
- langchain.com/blog/langchain-langgraph-1dot0 — Official 1.0 release: first stable major in the durable agent framework space, no breaking changes until 2.0, only notable change is deprecation of langgraph.prebuilt into langchain.agents (accessed 2026-09-04)
- docs.langchain.com/oss/python/versioning — Official version policy: LangGraph 1.0 designated LTS, ACTIVE until 2.0 then MAINTENANCE for at least 1 year; legacy LangGraph 0.4 in MAINTENANCE until December 2026 (accessed 2026-09-04)
- langchain.launchnotes.io/announcements/ann_NXVDQVGq… — Official naming change October 2025: LangGraph Platform renamed LangSmith Deployment, LangGraph Studio renamed LangSmith Studio; naming only, existing deployments unaffected (accessed 2026-09-04)
- langchain.com/about — Company-wide official figures: 35% of the Fortune 500, 1 billion+ open-source downloads, 1 billion+ events per day on LangSmith; HQ San Francisco with New York, Boston and Amsterdam offices; LangSmith positioned as the commercial layer over the open-source frameworks (accessed 2026-09-04)
- pub.towardsai.net/langgraph-checkpointing-is-not-fr… — Independent production postmortem (30/06/2026): checkpoint schema changes break in-flight threads with no migration system, interrupt behaviour undefined mid-fan-out in parallel subgraphs, every step writes a full blob synchronously in the hot path, no TTL, MemorySaver hides serialization bugs that PostgresSaver surfaces (accessed 2026-09-04)
- dev.to/elenarevicheva/three-langgraph-rewrites-what… — Independent practitioner account (17/06/2026): default overwrite reducer silently drops state, TypedDict validates nothing at runtime, SQLite checkpointer locks under concurrency, checkpoint and side effect are not transactionally linked so resume duplicates non-idempotent calls; 99.3% completion rate after hardening (accessed 2026-09-04)
- lifetideshub.com/langgraph-multi-agent-production-g… — Independent multi-agent production data across 847 runs April-June 2026: supervisor routing consumed ~40% of the token budget, no built-in termination guarantee so agent pairs ping-ponged 12 iterations, adding a state key stranded 17 in-flight workflows requiring manual database surgery (accessed 2026-09-04)
- bearplex.com/stack/langgraph-review — Independent agency review (15/06/2026) after 8+ production systems: explicit state and checkpoints called the decisive features, plus documented drawbacks — steeper learning curve than LangChain, TypeScript port lags Python, frequent releases occasionally break, deepest observability requires LangSmith (accessed 2026-09-04)
- neurals.ca/tech/langchain/langgraph/multi-agent — Independent reference (refreshed 27/05/2026) on the four documented multi-agent shapes — network, supervisor, swarm, hierarchical — the langgraph-supervisor-py and langgraph-swarm-py prebuilt packages, and handoffs persisted through the checkpointer (accessed 2026-09-04)
- truefoundry.com/blog/langgraph-pricing — Independent pricing breakdown: LangGraph framework free and MIT-licensed, hosted deployment formerly LangGraph Platform now part of LangSmith Deployment on the $39/seat Plus plan with one free small serverless deployment, further usage metered (accessed 2026-09-04)