Doubling Down on LangChain
CapitalG is proud to lead LangChain's $300M Series C at a $3B valuation, about a year after we first invested in their Series B, as LangChain builds the platform for every company to own its intelligence.
When we first invested in LangChain in 2025, most agents were still prototypes. Developers were building with LangChain's frameworks in huge numbers, but few companies were ready to let an agent work with customers or ship code on its own. That changed once the model generations released in late 2025 could work through a task for hours instead of minutes. Over the past year, agents at some of the largest companies in the world have moved into meaningful roles such as customer support, back-office work, and software engineering. A year ago, most enterprises still treated agents as promising science projects. Today the question for CIOs is how to build agents they can trust, run them at scale, and improve them every week.
LangChain calls this workflow “Agent Engineering”. Agents are non-deterministic, so getting one into production takes an iterative loop of testing, measuring, and refining, with product, engineering, data science, and AI all working together. Inside large enterprises, this is quickly becoming its own function, with dedicated teams and tools. In many ways, it feels like the agent development lifecycle (ADLC - build, test, deploy, monitor, iterate and govern) is taking a similar shape that the software development lifecycle took a generation ago, when source control, CI / CD and monitoring became standard. LangChain is building the leading platform that supports every stage across the emerging ADLC.
Owning your intelligence
Every company now has access to the same frontier models through the same APIs. That's great for adoption, but it also means the model on its own doesn't give any company an edge for long. So what makes an agent good at a specific job? For the most part, it's everything around the model: the harness that decides how the agent plans and uses tools, the context that tells it what matters for this business, the evaluations (evals) that define what good looks like, and the memory it builds up over time. As LangChain co-founder and CEO Harrison Chase put it in a post this summer, “a generic model [for an insurance company] knows the meaning of the word 'deductible'. But it does not know how to handle a specific type of claim, for this customer, in this jurisdiction, under this policy, with this evidence.”
LangChain's view is that the agents that matter most will be domain-specific, differentiated by the knowledge, data and user experience behind them. Some will be sold as products, and many more will run inside companies' own operations. Either way, the companies that control these systems will build lasting advantages.
LangChain's mission is to enable every company to own its intelligence, and it's a view we share.
In practice, owning their intelligence frees companies to choose the best model for each job, control their own data and infrastructure, and keep improving their agents, with the governance to run all of it securely at scale. Said differently, the 100th time someone uses an agent should be better than the first, and the company should keep what the agent learns along the way.
Owning those layers also makes it much easier to change models. The top spot on the leaderboards has changed hands several times this year, and the best open-weight models now trail the frontier by a few months at a fraction of the cost. Inference has also become a big enough line item that CFOs are asking how it's managed. Teams that can swap models without rebuilding their agents, and prove with evals that nothing regressed, get to use the best model the day it ships and an open-weight model wherever cost or control matters more.
Why we're doubling down
A year into our Series B investment, three beliefs give us the conviction to lead LangChain’s Series C.
1) The open source community keeps growing. LangChain began as Harrison's side project in late 2022, and its open source projects have since passed 1 billion downloads. At the Series B, LangChain's two core packages (LangChain and LangGraph) had ~90M monthly downloads combined. Today its open source projects, now including Deep Agents, see more than 350M downloads a month. The community matters because it's where new agent patterns get worked out first, from Deep Agents, LangChain's open source harness for long-running autonomous work, to LangGraph for production agents that need control. It's also how LangChain meets its enterprise customers, since many of the largest deployments started with a developer who already knew the frameworks.
2) LangSmith now runs, improves and governs agents in one place. A year ago, LangSmith was best known for tracing and evals. LangChain now organizes it around three pillars. Agent Runtime covers how agents get built and run, from the open source frameworks to Agent Server and Sandboxes for running agents securely in the cloud or in a company's own environment, plus fully managed options in LangSmith Managed Agents and Fleet. Agent Improvement covers observability and evals, and Fine-Tuning turns agent traces into training data for specialized models. Agent Governance gives companies one place to manage model budgets and fallbacks, protect sensitive data and control which tools agents can use, through LangChain's LLM and Tool Gateways. A piece we’re especially excited about is LangSmith Engine, LangChain's in-platform agent (i.e., an agent that improves other agents), which finds errors in production, writes the code or prompt fix and then watches for regressions. This is the feedback loop that makes an agent better with use, and LangSmith is in a good position to run it since it already sees every step agents take in production.
3) Enterprises want a neutral layer they control. The model labs and hyperscalers are all shipping their own agent tooling, and much of it is very good. An enterprise running hundreds of agents across several model providers and clouds still needs one place to build, evaluate, govern and improve all of them. What we hear most often from LangChain's customers is that they want that layer to be neutral and under their own control. LangSmith is framework-agnostic, works with any model, and runs in LangChain's cloud, self-hosted, or in the customer's own AWS, Azure or Google Cloud account. More than 7,000 organizations rely on LangSmith today, from startups like Harvey, Rippling and Listen Labs to enterprises like NVIDIA, Toyota and Salesforce. Clay uses it to monitor hundreds of millions of agent runs a month, and Workday uses it to run and govern agents across the company, which has cut the time to deploy a new agent from five months to under a day.
The team
LangChain founders Harrison Chase and Ankush Gola started the company in early 2023. Harrison has called each shift in agent development early. LangGraph shipped before most teams had an agent in production, and Deep Agents and Engine came out well before harnesses and self-improving agents were on most enterprise roadmaps. Ankush has built the engineering organization to match that pace, and this fall alone the team shipped Engine v2, Fine-Tuning, and BYOC on Azure and Google Cloud, with more coming at LangChain's release event on November 20. Every time we sit down with Harrison, Ankush and the team, we come away with a clearer view of where the market is going.
We believe LangChain is on its way to becoming the default platform for companies that want to own their intelligence. We're thrilled to lead LangChain's Series C alongside GIC, DFJ Growth, IVP and Sapphire Ventures and to deepen our partnership with Harrison, Ankush and the entire LangChain team.
