LangGraph

LangGraph: the engineer's route out of a stalled pilot

When the post mortem on a stalled pilot reads like an engineering incident report (runs that died halfway through, approvals bolted on with email, a demo that never survived concurrent users), the conversation tends to turn to LangGraph. LangChain's page for the product headlines it as an orchestration framework for agents that need to be reliable¹ and describes it as both an agent runtime and a low level orchestration layer². The documentation's definition is more precise: a low level framework and runtime for agents that are stateful and run for a long time³. On GitHub the tagline is shorter still, a promise to build resilient agents⁴, with the repository describing a low level framework for stateful agents⁵.

This site ranks LangGraph sixth. Of the seven contenders it gives an engineering team the most direct control over the failure modes that kill pilots in production, and its hosting options run from LangChain's cloud to a customer's own Kubernetes cluster. It ranks low for the buyer this site is written for because it is a developer framework: everything from the first node to authentication is the customer's code, and a vendor team that builds the agent for the customer is not publicly stated on the pages we opened.

Library, runtime and service: three things with one name

Buyers often meet LangGraph through a developer who has already used it, and the commercial picture needs untangling before procurement. The library is open source under the MIT License^6,7, and the FAQ says it is free to use⁸ and adds no overhead to the code it runs⁹. It does not dictate prompts or agent architecture¹⁰, and it can be used with no dependency on LangChain¹¹, even though LangChain Inc builds it¹². On GitHub, where engineers signal adoption with stars, the LangGraph repository had collected 44.7k by the time we looked¹³.

Running agents in production is a separate product. Since October 2025 the hosted runtime formerly sold as LangGraph Platform carries the LangSmith Deployment name¹⁴; LangChain describes it as the managed way to run those agents in production at scale¹⁵, and it accepts agents from frameworks other than LangGraph as well¹⁶. The overview page draws the family tree: LangChain supplies abstractions and integrations covering agent loops, tools and models¹⁷; LangSmith handles deployment, prompts, evaluation and tracing for any framework¹⁸; and LangGraph needs neither of them to function¹⁹.

The features that answer production failures

The reason engineers reach for LangGraph is that its core features map onto the ways agents break once real users arrive. Persistence writes the agent's state to checkpoints, and a run picks up again from the last one when something fails or a person pauses it²⁰. Human in the loop lets a run stop, have its state inspected or edited, and then carry on²¹, and the homepage frames the same idea as checks through which people steer and approve what an agent does²². Hand coded nodes and nodes driven by a model can sit together in one graph²³, which lets a team keep a deterministic step, such as a payment or a write to the ledger, outside the model's discretion. Memory covers both the state of a thread in the short term and longer term recall across sessions²⁴, and streaming emits tokens, state changes and other events while the graph executes^25,26.

The underlying model is a graph of shared state, nodes and edges²⁷. Two storage mechanisms do the work. A checkpointer records the graph state of one thread as successive checkpoints, which serves as memory for that conversation²⁸; a store holds whatever data the application chooses, kept apart from graph state, so it can travel across threads²⁹. For a sponsor this vocabulary matters less than the outcome: a pilot rebuilt on these primitives can survive a crash in step seven without starting again from step one.

Testing before anything ships

LangSmith Studio is presented as a free visual interface for building and testing agents³⁰. Pointed at an agent on a developer's own machine, it displays every step as the agent works through a task³¹, and when tracing is disabled in the agent's .env file the documentation says nothing leaves that machine³², a useful property when the pilot's test data is sensitive. The LangGraph CLI supplies the local development server Studio connects to³³. In production, a component called LangSmith Engine goes through the traces that LangGraph agents leave behind, spots faults and recommends how to correct them³⁴.

Evaluation is metered separately. Tuned Evaluators are charged only when one runs successfully³⁵, a Perceived Error evaluator costs 0.015 LSU per run³⁶, and the public beta of Tuned Evaluators is open only to Plus and Cloud Enterprise customers in the United States³⁷.

From repository to running deployment

Getting started is a developer task. The package installs with uv add langgraph³⁸, and the overview sends readers to the quickstart for a complete calculator agent³⁹. Moving to LangSmith Cloud starts with code in GitHub⁴⁰, and the guide accepts private repositories as readily as public ones⁴¹. A new deployment is created with the + New Deployment button⁴² and may take around 15 minutes to finish⁴³. When it finishes, the developer can open the live graph in Studio from the button at the upper right of the screen⁴⁴, take the endpoint address from the Deployment details page⁴⁵, and invoke the agent from Python through the LangGraph SDK⁴⁶.

The deployment service is not limited to LangGraph code. The documentation lists AutoGen, CrewAI, Strands and the Claude Agent SDK among the frameworks whose agents it can run⁴⁷, and it explains how an agent written with Google's Agent Development Kit can be wrapped and deployed as a LangGraph⁴⁸. For a company with several stalled pilots on different frameworks, that is a way to consolidate operations without rewriting everything first.

Connecting agents to the rest of the estate

LangGraph's own pages describe integration through protocols and triggers more than through a connector catalog. On the deployment side, the agent registry speaks Agent Protocol, MCP and A2A out of the box⁴⁹; scheduled runs are handled by crons⁵⁰; outbound webhooks let an agent set other systems in motion⁵¹; and the pricing page lists the option of publishing an agent as an MCP server⁵². Model and tool integrations come through LangChain¹⁷, and the homepage says agents can be built and shipped quickly with any model provider⁵³. A count of prebuilt connectors to business applications comparable to the figures other vendors publish is not publicly stated on the pages we opened, which means the customer's engineers write or source each connection.

Where the agents and their data can run

Hosting flexibility is one of LangGraph's real strengths for a pilot that stalled over data control. According to the deployment docs, agents can be placed in LangChain's cloud, in a bring your own cloud setup, or on hardware the customer operates⁵⁴, and the pricing page lists the hosting choices as Cloud, Hybrid or Self Hosted⁵⁵. In LangSmith Cloud, LangChain runs the control plane, the data plane and the databases⁵⁶. The hybrid arrangement splits responsibility, with the control plane staying at LangChain and the Agent Servers, together with their data plane, running at the customer⁵⁷. Under self hosting both the Agent Servers and the control plane move onto the customer's infrastructure⁵⁸, and this is the option the documentation points to where full residency or an air gap is required⁵⁹. The pricing table records that self hosted data lives in the customer's VPC⁶⁰, and LangSmith itself can be installed on a Kubernetes cluster the customer manages⁶¹.

Two details deserve attention in a hybrid design. Agent Servers send traces to whatever endpoint the LANGSMITH_ENDPOINT variable names⁶², and if that variable is left out, traces go to the default GCP US endpoint⁶³. Cloud customers choose between LangChain's US or EU regions⁶⁴. Self hosting is not available on every plan: it is an add on to Enterprise⁶⁵, it keeps trace and operational data inside the customer's environment⁶⁶, and a license key for a self hosted trial requires contacting sales⁶⁷.

A compliance officer asking about training gets a direct answer in the pricing FAQ: customer data is not used to train models⁶⁸, and every trace, prompt and output remains visible only within the customer's own organization⁶⁹. Retention depends on the trace tier, two weeks for base traces⁷⁰ against roughly six months, 180 days, for extended ones⁷¹, a difference an audit team will want settled before go live.

Security and access, largely in the customer's hands

This is the area where a security reviewer will have the most questions. LangSmith offers an auth framework flexible enough to plug into most existing schemes⁷², but the logic is the customer's: its developers write the @auth.authenticate handler that verifies identity and the @auth.on handlers that grant or refuse access⁷³. Custom auth works on every LangSmith plan⁷⁴, and authorization checks each user's roles and privileges resource by resource⁷⁵. The trade off is stated plainly: there is no default authentication, so the customer controls both authentication and authorization⁷⁶. When switched on, the same handler will also honor LangSmith API keys next to the custom scheme⁷⁷, and the platform page lists custom auth and access control as a feature⁷⁸.

Organization level controls are tiered. The security table on the pricing page names two Organization Roles, Admin and User⁷⁹, and lists role based access control⁸⁰; SSO below Enterprise is limited to Google and GitHub⁸¹, while Enterprise adds custom SSO, ABAC and RBAC⁸², custom seats and workspaces⁸³, and an Infosec Review as part of procurement⁸⁴. The agent registry manages agents centrally with versioning⁸⁵, and assistants can be shared, modified and reused⁸⁶. Agent Authorization is still in beta⁸⁷.

For certifications, a reviewer is pointed by the homepage footer toward a Trust Center⁸⁸, yet when we opened trust.langchain.com our saved copy contained nothing beyond the page title⁸⁹. SOC 2, ISO 27001, HIPAA and GDPR certifications are not publicly stated on the pages we opened. A reviewer will need to obtain them through the Infosec Review route.

Limits that matter in production

A few documented limits translate directly into production risk. MemorySaver, the in memory checkpointer, does not survive a restart⁹⁰, so a production deployment needs a persistent checkpointer. Checkpoints that grow without bound are listed as a known issue⁹¹, and persistent checkpoints on the deployment service accept payloads of up to 25 MB⁹². Sandboxes are capped at ten⁹³. For a production service that needs guaranteed response times, note that the pricing page attaches a support SLA to the Enterprise plan⁹⁴.

What it costs

LangSmith pricing is published. A single developer can start on the free tier, which charges nothing per seat⁹⁵ and covers 5k base traces monthly before metered billing begins⁹⁶. The team tier, Plus, is $39 a seat monthly⁹⁷ with 10k base traces included each month⁹⁸, and there is no ceiling on the number of seats bought at that price⁹⁹. The FAQ counts as a seat each separate person in the organization, people who have only been invited among them¹⁰⁰. That tier also bundles a small serverless deployment at no extra charge, enough to put one rescued pilot online¹⁰¹. Enterprise has no list price¹⁰² and is billed a year at a time, up front¹⁰³; on the two lower tiers, by contrast, trace usage is charged each month after the fact¹⁰⁴.

Consumption beyond the plans is counted in a unit LangChain calls the LSU, priced at one dollar¹⁰⁵. Each vCPU hour of runtime on a deployment uses 0.0675 LSU¹⁰⁶; a cloud deployment's bill follows the database and compute capacity it asks for¹⁰⁷; and the sandbox allowance is 8 LSU per month¹⁰⁸. Startups can apply for up to $10,000 in credits¹⁰⁹.

Who uses it

The buyers LangChain names are developers. The homepage speaks to developers first, saying they rely on LangGraph for reliable agents¹¹⁰ and pitching the fine grained control it gives over how an agent is built¹¹¹; it also claims the trust of firms at the front of agent development¹¹². Named users on GitHub include Elastic, Replit and Klarna¹¹³, and the homepage quotes Ally, whose technology labs use LangGraph, as part of Ally's exploration of generative AI, for agentic workflows involving several actors¹¹⁴. Plan descriptions keep the engineering focus. The free tier is meant for one person taking first steps¹¹⁵, Plus for teams that build and deploy agents¹¹⁶, and Enterprise for teams whose hosting, support and security requirements go beyond the standard¹¹⁷, which is where a regulated production rollout would land. LangSmith Deployment is pitched at helping top teams ship great agents¹¹⁸.

Verdict for a team with a stalled pilot

LangGraph ranks sixth. For an engineering led organization whose pilot failed on reliability, it is a serious option: durable checkpoints²⁰, pause and edit human review²¹, deterministic and model driven steps in one graph²³, and a hosting range that reaches self hosted and air gapped deployments^58,59. Its open source license⁶ keeps the core free of vendor lock in.

It ranks below the five platforms ahead of it for the buyer this site addresses because the customer writes the agent, the connectors and the authentication⁷⁶, certifications are not publicly stated on the pages we opened, and self hosting sits behind an Enterprise add on and a sales conversation^65,67. A sponsor whose pilot stalled for lack of engineering capacity will not find that capacity here; a sponsor with a strong platform team may find LangGraph the most controllable foundation on the list.