LangChain

LangChain is the company behind the LangChain open source framework, LangGraph, Deep Agents, and the commercial LangSmith platform. Cofounded by Harrison Chase and Ankush Gola, it builds software for connecting language models to data and tools, controlling agent execution, and understanding failures. The company and its namesake framework have different scopes: LangChain supplies agent building blocks, LangGraph provides stateful orchestration, Deep Agents adds a fuller harness for extended tasks, and LangSmith supports tracing, evaluation, and deployment across frameworks.[1][9][10][11]

This product stack grew from a single Python package released in late 2022. Its early prompt-handling and integration work supplies context for the company's later emphasis on debugging and explicit control over agents.[1][3][4]

In a September 2023 interview, Chase recalled building a question-answering bot over Notion and Slack at a Robust Intelligence hackathon with Will, who later joined LangChain, and two other colleagues. He could not precisely reconstruct the sequence of events. He distinguished that experiment from the framework's immediate inspiration: conversations at August and September 2022 meetups revealed common components among language-model applications. He thought it would be “a cool side project to factor out some of those common abstractions”.[2]

An early pull request merged on October 17, 2022 defined a prompt schema with a template and declared input variables. It validated the template through stricter Python string formatting that rejected extra keyword arguments and positional arguments. These safeguards let developers represent prompts explicitly and catch mismatches between templates and inputs.[3]

The package expanded through integrations with models and vector databases, alongside higher-level templates for retrieval-augmented generation, SQL question answering, and extraction. Chase and Gola formed the company in early 2023. Supporting multiple model providers remained a deliberate choice: Chase explained that developers needed to select models and change that decision as alternatives emerged, while acknowledging that prompts did not necessarily transfer between models.[4][1][2]ⁿ¹

LangSmith entered closed beta on July 18, 2023 and became generally available on February 15, 2024. It exposed intermediate prompts, raw outputs, call sequences, latency, and token usage, making failures within multi-step applications inspectable. Developers could turn application traces into datasets of inputs and reference outputs, then compare results after changing a prompt, model, or retrieval strategy. Tracing shows what happened; evaluation tests the effect of a change.[5][6]

A concrete workflow shaped its debugging interface. According to the beta announcement, teams copied failed examples from logs into the OpenAI playground to adjust prompts. LangSmith added an edit-and-rerun route from a logged example to its own playground. Chase later explained that the company deliberately kept LangSmith separate from the LangChain framework so developers using other stacks could use its observability and evaluation tools.[5][4]

In his three-year retrospective, Chase acknowledged summer 2023 feedback about hidden prompts, package bloat, dependency conflicts, outdated documentation, and insufficient control. He described the original framework's high-level interfaces as easy to start with but restrictive when developers needed production customization. The team began developing LangGraph that summer, emphasizing explicit orchestration and an execution runtime.[4]

The January 17, 2024 LangGraph announcement explained cycles through retrieval-augmented generation, or RAG: an application retrieves documents to help a model answer a question. If the first retrieval returns poor material, an agent can assess it, refine the query, and retrieve again. LangGraph represents such flows as graphs with shared state, nodes that update it, and fixed or conditional edges—including a route back to retrieval. Its launch described it as built on LangChain; current documentation says LangGraph can be used without LangChain.[7][10]

In his September 2025 runtime design account, Nuno Campos illustrated the cost of restarting with a hypothetical ten-minute agent run that fails in minute nine. LangGraph's checkpoints save computation state so recovery can avoid repeating the entire run. Streaming exposes useful progress while execution continues, and interruption and resumption allow a person to approve an action or inspect and modify state without keeping the process running while it waits.[8][10]

LangGraph separates its PregelLoop execution runtime from developer-facing StateGraph and functional APIs, allowing those layers to evolve independently. Parallel nodes execute with isolated state copies; their updates are then applied in a deterministic order. Its documentation credits Pregel and Apache Beam as inspirations and NetworkX for its public interface. The task queue required for deployment was handled through LangGraph Platform rather than the Python library.[8][10]

The company released LangChain and LangGraph 1.0 on October 20, 2025. Chase described LangChain's redesign around the tool-calling loop, middleware, and LangGraph's runtime. The rewrite involved major breaking changes, while the earlier package remained available as langchain-classic. The current LangChain overview summarizes its architecture as “Agent = Model + Harness.” Its create_agent interface composes a model, tools, a prompt, and middleware—the surrounding software that shapes the model loop's behavior.[14][4][9]

The same October 2025 announcement brought deployment, formerly LangGraph Platform, into LangSmith, alongside observability and evaluation. It also introduced an Insights Agent for categorizing behavior patterns and a no-code Agent Builder in private preview. This expanded the commercial platform from inspecting and testing agents to supporting their deployment and creation, while retaining support for agents built with other frameworks.[14][4]

Before the October 2025 releases, LangChain's July 2025 context-engineering synthesis organized practices from agent products, papers, and practitioners into four operations: write, select, compress, and isolate. A scratchpad illustrates the mechanism: an agent saves information outside the model's context window, then retrieves selected portions when needed. Summarization reduces accumulated history, while separate subagent contexts isolate working material. LangGraph's state and memory mechanisms support these operations; LangSmith supplies traces and evaluations for testing their effects.[12]

Deep Agents packages more of this machinery around the tool-calling loop. It is a standalone library using LangChain building blocks and LangGraph's runtime, with filesystem tools, persistent memory, history summarization, offloading of large results, and subagents running in isolated contexts. Planning and skills are optional capabilities, and human approval can be enabled for selected tool calls. These mechanisms manage information and execution across extended tasks.[11]

In February 2026, Chase characterized the company's progression as chaining, orchestration, and then fuller agent harnesses. He argued that frameworks must evolve with models, while acknowledging a limit: “If it’s a simple LLM request, adding a framework may be too heavy handed.” LangSmith remained independent of that choice, supporting other frameworks, OpenTelemetry traces, and agents built without a framework.[13]

LangChain announced a $25 million Series A led by Sequoia Capital with LangSmith's general availability on February 15, 2024. On October 20, 2025, it announced $125 million raised at a $1.25 billion valuation, led by IVP, with existing investors Sequoia, Benchmark, and Amplify and new investors CapitalG and Sapphire Ventures. TechCrunch also reported a $10 million Benchmark seed round in April 2023 and placed the Series A raise a week later; the February 2024 announcement therefore should not be read as its closing date.[6][14][15]

At LangSmith's February 2024 general availability, the company reported over 80,000 signups, over 5,000 monthly active teams, and over 40 million traces logged in January. In October 2025, it reported 90 million combined monthly downloads for LangChain and LangGraph and said 35% of the Fortune 500 used its services. These are company-reported adoption and activity figures, rather than counts of paying customers or revenue. LangChain is headquartered in San Francisco, with offices in New York, Boston, and Amsterdam.[6][14][1]