LangChain

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

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

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

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; 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

The frameworks run inside the customer's application with no hosted service or user layer. The documented controls hold tools to least privilege, through Deep Agents filesystem permission rules and a security policy on scoping agent credentials and sandboxing. No supported identity providers or provisioning standards are named for the frameworks, since identity and authorization are documented as coming from LangSmith's auth layer, and no attestation is published for the open source packages.

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

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. The documented tracing, monitoring and dashboards require a LangSmith account, and the frameworks document no audit log separate from runtime traces.

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

LangGraph documents two persistence systems with a stated scope and lifetime: 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

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

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. Each does its own job without the other. The create_agent harness and the tutorial agents are building blocks for assembly rather than finished agents.

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

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 documented as LangSmith Deployment features, a separate product.

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

The developer initializes a model from any supported provider with init_chat_model, and the provider catalog spans OpenAI, Anthropic, Google and many others, so model choice is the customer's.

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

The product is itself a documented 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

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

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 needing its own account and key, so they are products sold beside the framework rather than shipped with it. Deep Agents sandboxes run code rather than operating a browser or desktop.

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