LangChain vs Vellum (2026)

Agents act in outside systems through custom tools, callable functions the model invokes to fetch data, query databases and take actions, plus tools from MCP servers through the langchain-mcp-adapters library and a large provider integration catalog.

SourceLangChain, docs.langchain.com/oss/python/langchain/tools and /langchain/mcpread 2026-09-21

The Agent Node calls tools with automatic function-calling schema generation and loop logic, drawing tools from custom code, subworkflows and API calls, and function calling with chat models is supported, including parallelized function calls. The API Node makes authenticated HTTP requests to any endpoint and the Code Execution Node runs custom Python or TypeScript, and examples cover a Zapier and Airtable integration and a Datadog export. Agents can act in outside systems through custom tools.

Sourcedocs.vellum.ai Agent Node, API Node, Code Execution Node and function-calling pages, 31 August read, re-gradedread 2026-09-29

LangGraph is a low level orchestration runtime that mixes deterministic, hand coded steps with LLM driven agent steps in one graph, with durable execution, subgraphs and multi agent patterns, so fixed nodes and autonomous agent steps run in the same flow.

SourceLangChain, docs.langchain.com/oss/python/langgraph/overviewread 2026-09-21

Fifteen node types cover an Agent Node with automatic schema handling and loop logic, prompt and prompt-deployment nodes, templating, a search node, an API node, a code execution node running Python or TypeScript, and subworkflow, map, guardrail, conditional, merge, final output, error and note nodes.

Node adornments apply retry and try semantics to existing nodes, and long-running workflows and batching executions are supported. Multi-agent composition works through subworkflows, with worked examples including multi-agent content creation and LLMs debating each other, and six common architectures cover RAG, escalation to a human, prompt retry, PDF summarization and fallback models.

Sourcedocs.vellum.ai workflows nodes overview and per-node pages, node-adornments, advanced long-running-workflows and batching-executions, common-architectures and examples pagesread 2026-08-31

Every run starts when the customer's own code invokes or streams the agent, so every channel is one the customer wires up, and the open source frameworks ship no schedule, webhook or event source that wakes an agent on its own. Webhooks and cron are LangSmith Deployment features, a separate product.

SourceLangChain, docs.langchain.com/oss/python/deepagents/going-to-productionread 2026-09-21

Workflows are invoked programmatically through a deployed API and called from application code, with batching executions for volume and long-running workflows for extended jobs. The webhook integration pushes execution data outward to customer systems. There is no scheduled, cron, event-driven or chat-channel trigger.

Sourcedocs.vellum.ai workflows api-integration, advanced batching-executions and long-running-workflows, monitoring webhooks and execution-urls pages, and Product navigation indexread 2026-08-31

The frameworks ship a retrieval layer in which documents are loaded, split, embedded and indexed in a vector store, then queried by similarity through a retriever, so new customer knowledge enters a persistent index without retraining. The vector store backend is one the customer chooses from the integration catalog.

SourceLangChain, docs.langchain.com/oss/python/langchain/knowledge-baseread 2026-09-21

Documents are uploaded into a Document Index, with a Search API integration and metadata filtering for scoped retrieval. A dedicated Search Node searches against a Document Index for RAG. A RAG system is one of the common architectures, with worked examples covering a basic RAG chatbot, a RAG chatbot with Cohere rerank and a building-a-RAG-chatbot tutorial, and RAG pipelines can be evaluated under Evaluation and Test Suites. Documents persist as an index between runs.

Sourcedocs.vellum.ai documents section, workflows nodes search-node, common-architectures rag-system, workflow examples and evaluation/evaluating-rag-pipelines pagesread 2026-08-31

LangGraph has two persistence systems. Checkpointers persist a thread's state as short-term, thread-scoped memory, and stores persist long-term memory across threads, namespaced by user or any other key, which agents read and write from their tools.

SourceLangChain, docs.langchain.com/oss/python/langgraph/persistence and /langchain/long-term-memoryread 2026-09-21

A Document Index persists uploaded material between runs and is queryable through a Search Node or Search API. Long-running workflows handle executions that extend beyond a single request, and deployment lifecycle management, environments and release tags persist configuration and version history across releases. There is no memory module, conversation store, session identity or cross-execution context capability, and workflows are invoked per execution through the deployed API.

Sourcedocs.vellum.ai workflows advanced long-running-workflows, documents section, deployments section and Product navigation indexread 2026-08-31

Human-in-the-loop middleware checks each tool call against a configurable per-tool policy, halts on an interrupt with state saved, and lets a person approve, edit or reject the call before execution resumes. The built-in guardrail middleware adds PII detection and model and tool call limits.

SourceLangChain, docs.langchain.com/oss/python/langchain/human-in-the-loop and /langchain/guardrailsread 2026-09-21

A Guardrail Node runs an inline evaluation against a pre-defined Metric inside a workflow, Error and Conditional nodes stop or branch execution, Retry and Try adornments handle failures, and release reviews come before a change is promoted to a live deployment.

Vellum also has role-based access control and an Escalation to a Human architecture that automatically routes sensitive or complex messages to human operators. These are runtime constraints, a design-time release step and workflow-initiated escalation, and there is no approval node or pause for a person to approve an agent's action before it executes.

Sourcedocs.vellum.ai documentation index, Escalation to a Human, Guardrail Node and release reviews pagesread 2026-09-29

The frameworks run inside the customer's application with no hosted service or user layer. Tools are held to least privilege through Deep Agents filesystem permission rules and a security policy on scoping agent credentials and sandboxing. Identity and authorization come from LangSmith's auth layer, so no identity providers or provisioning standards are named for the frameworks themselves, and no attestation is published for the open source packages.

SourceLangChain, docs.langchain.com/oss/python/deepagents/permissions and /security-policyread 2026-09-21

Vellum maintains SOC 2 Type 2 compliance and is HIPAA compliant, with security practices regularly audited to meet industry standards and healthcare data protection requirements. All data stored in Vellum, including documents in Document Indexes, is encrypted with AES-256 GCM in transit and at rest.

Vellum does not send interactions or feedback to LLM providers for training, and Completion Actuals submitted through the feedback API are stored for the customer's own quality monitoring rather than used to train or fine-tune models.

Controls include role-based access control, HMAC authentication for verifying request origin, static IPs for customer-side allowlisting, organization access management and configurable data retention policies.

Sourcedocs.vellum.ai security section covering data-privacy-and-storage, rbac-permissions, hmac-authentication and static-ips, and organizations manage-access and pagesread 2026-08-31

The open source frameworks expose step-by-step state history through checkpoints that can be replayed or forked, and stream each step as it runs, so tool calls and outputs can be inspected per step. Tracing, monitoring and dashboards require a LangSmith account, and the frameworks have no audit log separate from runtime traces.

SourceLangChain, docs.langchain.com/oss/python/langgraph/observability and /langgraph/use-time-travelread 2026-09-21

Production observability comes with monitoring of production trends, tracking of workflow execution costs, and execution URLs that address individual runs. A Datadog integration and a webhook integration export execution data to the customer's own monitoring and alerting systems, and online evaluations score production traffic on defined metrics. Deployment lifecycle management, environments and release tags tie each execution to a specific released version, and data retention policies are set at organization level.

Sourcedocs.vellum.ai deployments observability and deployment-lifecycle-management, monitoring section covering production-trends, execution-cost-tracking, datadog, webhooks and execution-urls, evaluation online-evaluations, and organizations pagesread 2026-08-31

The frameworks install as packages into the customer's own application and run on the customer's infrastructure, with state persisted to checkpointer backends the customer operates, such as SQLite and Postgres, so data storage stays where the customer puts it. Managed cloud, hybrid and self hosted server options belong to LangSmith Deployment, a separate product.

SourceLangChain, docs.langchain.com/oss/python/langgraph/checkpointers and /langgraph/deployread 2026-09-21

Self-hosting is supported, with its own getting-started introduction. Static IPs, available for security and for advanced workflow behavior, give fixed egress addresses for customer-side allowlisting. Data retention policies are configurable at organization level, and a data privacy and storage policy sets out how data is held. Managed deployment covers deployment lifecycle management, environments and release tags for promoting versions.

Sourcedocs.vellum.ai self-hosting introduction, security static-ips and data-privacy-and-storage, workflows advanced static-ips, organizations and deployments pagesread 2026-08-31

LangChain publishes two complete open source agents built on Deep Agents: Deep Agents Code, a terminal coding agent with persistent memory, skills and approval controls, and OpenWiki, a CLI that writes and maintains a Markdown wiki as durable context for coding agents. The create_agent harness and the tutorial agents are building blocks to assemble, not finished agents.

SourceLangChain, docs.langchain.com/oss/deepagents/code/overview and /oss/openwiki/overviewread 2026-09-21

Twelve named workflow examples cover prompt chaining, a basic RAG chatbot, RAG with Cohere rerank, a customer support bot, PDF to CSV, summarizing images of websites, parallelized function calls, conference attendee lookup, multi-agent content creation, LLMs debating each other, a Zapier and Airtable integration and automating PR reviews.

Six common architectures cover RAG systems, escalation to a human, prompt retry logic, PDF content summarization and fallback models. Out-of-the-box metrics are supplied by Vellum and reusable across test suites, and subworkflows let customers make their own work reusable. Vellum does not say whether an example can be imported into a workspace in one action.

Sourcedocs.vellum.ai workflow examples overview, common-architectures, metrics out-of-the-box-metrics and reusing-metrics, and workflows nodes subworkflow-node pagesread 2026-08-31

The developer initializes a model from any supported provider with init_chat_model, and the provider catalog spans OpenAI, Anthropic, Google and many others.

SourceLangChain, docs.langchain.com/oss/python/langchain/modelsread 2026-09-21

Custom Models bring models beyond the built-in roster, and prompt engineering covers comparing prompts across models. Fallback models are one of Vellum's common workflow architectures, and prompt caching and multimodality are supported. Model choice is tied to measurement through test suites and metrics, so switching providers can be evaluated quantitatively rather than by inspection. Prompt deployment nodes execute deployed prompts within workflows, and release tags and environments govern which prompt version, and therefore which model configuration, is live.

Sourcedocs.vellum.ai prompts custom-models, prompt-engineering, prompt-caching and multimodality, workflows common-architectures fallback-models, and deployments pagesread 2026-08-31

LangChain is itself an SDK, published for Python and TypeScript with a full API reference, and is called directly from the customer's own code, services and CI.

SourceLangChain, docs.langchain.com/oss/python/reference/overview and /langchain/overviewread 2026-09-21

Vellum offers SDKs, APIs and developer tools. Workflows deploy to an API that application code can call, and a Search API reaches the Document Index independently of workflows. Webhook integration exports execution data, and execution URLs address individual runs. A Code Execution Node runs customer Python or TypeScript within a workflow, and HMAC authentication verifies request origin for callbacks.

Sourcedocs.vellum.ai developers overview, workflows api-integration, documents api-integration, monitoring webhooks and execution-urls, workflows nodes code-execution-node and security hmac-authentication pagesread 2026-08-31

The open source agentevals package scores the customer's agent trajectory, by deterministic trajectory match in four modes or by an LLM judge, and a fake chat model lets unit tests script exact responses as fixtures, so a change can be evaluated before release without LangSmith. LangSmith datasets and online evaluations sit in LangSmith, a separate product.

SourceLangChain, docs.langchain.com/oss/python/langchain/test/evals and /langchain/test/unit-testingread 2026-09-21

Quantitative evaluation runs through Test Suites and Metrics, which Vellum presents as the answer to the regressions that prompt changes, parameter adjustments and model switches make likely. Metrics come out-of-the-box or custom and are reusable across test suites.

Online evaluations extend scoring to production, and RAG pipelines have their own evaluation. A Guardrail Node runs an inline evaluation using a pre-defined Metric within a workflow. Experimentation and a prompt playground support side-by-side comparison across models, and Retry and Try node adornments handle failure paths during development.

Sourcedocs.vellum.ai evaluation section covering quantitative-evaluation, online-evaluations and evaluating-rag-pipelines, metrics section, and workflows nodes guardrail-node and node-adornments pagesread 2026-08-31

The frameworks ship no browser, desktop or computer control of their own. Browser tools reach third party engines such as Browserbase and Bedrock AgentCore Browser, each sold separately with its own account and key. Deep Agents sandboxes run code and do not operate a browser or desktop.

SourceLangChain, docs.langchain.com/oss/python/integrations/providers/browserbaseread 2026-09-21

None of the fifteen node types provides browser, desktop or computer-use capability. The external-action nodes are an API Node making HTTP requests to an endpoint, a Code Execution Node running custom Python or TypeScript and a Search Node querying a Document Index, with tool calling handled by the Agent Node through automatic schema generation. One example summarizes images of websites, applying retrieval and vision to fetched material. There is no browser automation, session control, form filling or screen interaction.

Sourcedocs.vellum.ai workflows nodes overview and per-node pages, workflow examples summarize-images-of-websites, and Product navigation indexread 2026-08-31