LangChain spent the week of 17 November turning a funding announcement into a product story. On 13 November the company said it had raised $300 million at a $3 billion valuation. By Friday 20 November, at a release event led by co-founder and chief executive Harrison Chase, it had shown four new LangSmith capabilities that explain what the money is for: a rebuilt Agent Platform, a Tool Gateway, coding-agent ROI reporting, and new controls over sensitive data in traces. The funding is the headline. The platform rebuild is the news that matters to anyone already running agents in production.
What did LangChain announce this week?
Five announcements in eight days, each building on the last. The Series C set the frame; the rest filled it in.
| Date | Announcement | Status |
|---|---|---|
| 13 Nov 2026 | $300M Series C at a $3B valuation, led by CapitalG | Announced |
| 17 Nov 2026 | LangSmith rebuilt around the agent; three new trace-data controls | Announced; controls on Enterprise plans |
| 18 Nov 2026 | LangSmith for Coding Agents: spend tied to pull requests | Public beta |
| 19 Nov 2026 | LangSmith Tool Gateway: governed tool access | Public beta |
| 20 Nov 2026 | “The Future of Agents” release event with Harrison Chase | Held; recording published |
LangChain summed the week up itself in a post on X on 20 November, listing the four items: the new LangSmith Agent Platform, new controls for protecting sensitive data in traces, LangSmith for Coding Agents, and LangSmith Tool Gateway.
https://twitter.com/LangChain/status/2123838785960198512
The round itself is worth stating plainly. CapitalG led the $300 million Series C, with new investors GIC and DFJ Growth and existing investors IVP and Sapphire Ventures participating, alongside Salesforce Ventures, Insight Venture Partners, QuantumLight and MongoDB Ventures, with Sequoia, Amplify and Benchmark doubling down. The valuation is up from about $1.25 billion roughly a year earlier. It was also a big funding week for AI more broadly: Anthropic’s shares began trading the same week, closing their first day at $97.17 after opening at $111.50.
What is the LangSmith Agent Platform?
LangSmith is LangChain’s platform for running, improving and governing agents, and more than 7,000 organisations use it, according to the company. Until this month it was organised around runs and traces — the record of what happened inside a single call. That worked when an agent was one experiment. It broke down once companies ran dozens of agents across teams, because tracing, monitoring and deployments lived in different places and platform teams could not answer basic questions: which agents exist, who owns each one, what do they cost, and how are they performing?
The rebuild, announced on 17 November, makes the agent itself the core object. Every registered agent now has its own record with three properties: an owner, a named person responsible for it; an ID, a globally unique identifier that does not change when the agent is renamed, moved between environments or released in a new version; and a type, which determines what its dashboard shows, which data it connects to and how it is evaluated. Coding agents are the first specialised type.
Around that record sit the three pillars LangChain has used since the Series C announcement: Agent Runtime, the architecture and infrastructure to run agents reliably; Agent Improvement, testing, monitoring and evaluation, including the LangSmith Engine, an in-platform agent that finds errors, writes code and prompt fixes and watches production traffic for regressions; and Agent Governance, centralised control over how agents use models, tools and data.
The framing matters because of what LangChain is selling. As Harrison Chase put it in the Series C post:
“Companies should be free to choose the best models for their needs, control their data and infrastructure, and continuously improve their agents, while maintaining the governance needed to operate securely at scale.”
That is a pitch aimed squarely at enterprises that do not want to be locked to one model provider. LangSmith is framework-agnostic by design: teams can connect agents built on other platforms to improve and govern them without changing their other workflows. The customer figures LangChain cites are its own — Clay monitoring hundreds of millions of agent runs a month, Workday cutting the time to deploy a new agent from five months to under a day — and no independent verification of those figures has been published.
What does Tool Gateway actually do?
Tool Gateway is the most concrete of the four announcements, and the one most teams will touch first. It entered public beta on 19 November on the LangSmith Agent Platform. It is a managed tool layer: instead of each agent carrying its own credentials and OAuth flows, agents and MCP clients call one kit URL, and the gateway resolves identity, attaches the right credentials, applies policy, and logs the call.
The problem it solves is real and easy to picture. LangChain describes a large US retailer building agents for store operations that has built more than 50 MCP servers exposing over a thousand tools. One agent, a store assistant, should answer questions using each associate’s individual credentials. Another, an inventory agent, runs at 3 a.m. every day on its own service credentials. When credentials are hard-coded into agents, nothing deterministic stops an agent from using the wrong credential or an unapproved tool, and security has no end-to-end record to audit.
Tool Gateway organises the answer into three parts. Access: tools are registered once, either as any MCP server by URL or as a built-in server hosted by LangSmith for Google, Microsoft, Salesforce, GitHub, Linear and LinkedIn, and bundled into kits — curated sets of tools published at a single URL, so an agent only ever sees the tools in its kit. A search mode lets an agent with hundreds of tools load only the ones relevant to the current task. Control: admins set deny policies at user, service-key or org level, and approval policies that ask a person to confirm a sensitive call — writing to a data warehouse, for example — before it runs. Visibility: every tool call is traced in LangSmith with its server, user, outcome and latency, and on Enterprise plans an audit log answers the question “what tools did this agent use at 3 a.m. last night, and which credentials did it use?”
Identity is the part worth reading carefully, because it is where most hand-rolled agent setups fail. Each call acts under one of two identities. Acting as the user applies that person’s own credentials, identified through Okta or Entra ID; if the user has not connected yet, the call returns a sign-in link, waits for them to authenticate, then continues automatically. Acting as the agent uses credentials tied to the agent itself, either through agent authentication or a shared profile a workspace admin connects once. The agent itself never handles the token.
Two limits to note. The beta is available on LangSmith Plus and Enterprise plans, not the free tier, and on LangSmith Cloud it covers the GCP US, EU and APAC regions and AWS US — Azure US is excluded, according to Unite.AI’s report of the launch. There is also no general-availability date yet. The companion control, LLM Gateway, already governs model calls with spend limits, fallbacks, rate limits and redaction of personal data and secrets; Tool Gateway is its counterpart for tool calls.
How does LangSmith measure a coding agent’s ROI?
Coding is the largest category of enterprise AI application spend, reaching $34 billion in 2026, according to Menlo Ventures figures cited in LangChain’s announcement. Yet most engineering leaders still cannot say what that spend buys. Usage dashboards show tokens consumed; they do not show what the tokens produced.
LangSmith for Coding Agents, in public beta since 18 November, answers with a different unit: the pull request. Spend in each trace is automatically attributed to the pull requests the agent contributed to, including the messy cases — one thread opening several PRs, or several threads finishing one PR. A dashboard aggregates cost across every coding agent a team traces, breaks spend down by PR state (open, closed or merged), and tracks the rest separately as unattributed spend: reviews and investigations that never end in a PR but may still be useful work. It also shows work type (features, bug fixes, investigation, code review), which skills agents call, the most expensive PRs, and how cost scales with conversation length.
The practical uses LangChain suggests are coaching and comparison. Sorting by attributed cost shows which PRs were expensive to produce and why — a thread that kept growing, a frontier model used on a small fix. Long threads are a cost driver because every new turn carries the entire context, so starting a fresh thread per task is the cheap habit to teach. And when a team switches models, cost per merged PR before and after the change shows whether the results justify the price.
It works with Claude Code, OpenAI Codex, Cursor, GitHub Copilot in VS Code and Deep Agents, and setup is three steps: install the LangSmith plugin for each coding agent, connect GitHub, and open the coding-agent overview page. LangChain says its own coding agent, Robin, is part of why it built this: its spend kept climbing, which also led the company to build a model router. That is a candid origin story, and it is also a reminder that the metric is new — no independent team has yet published results from using it.
What do the new trace controls protect — and what do they not?
The fourth announcement, published on 17 November, adds three controls for sensitive data in LangSmith traces, available on Enterprise plans. First, admins can restrict who can see trace inputs and outputs, so engineers can debug with run trees, latency and cost statistics without seeing customer messages and model responses. Second, users can be organised into groups with team-by-team permissions layered on top of existing role-based and attribute-based access controls. Third, admins can set trace-level policies using root run metadata — for example, giving support agents access only to traces for the customers they support, using customer IDs, or developers access only to their own coding-agent traces.
Here is the honest caveat, and it matters. These are access controls, not redaction. Full traces are still stored in LangSmith; the controls decide who can see which parts. Hiding content from a teammate is different from removing it before upload, as independent analysis of the announcement points out. The pre-upload side is a separate set of features in LangChain’s SDK: hiding inputs and outputs entirely, transforming metadata, and an anonymiser that redacts recognised secret formats before traces leave your systems. Teams handling regulated data need both halves — what never leaves, and who can see what arrived — and should not mistake one for the other.
What is still unproven?
A week of announcements is not a week of evidence. Nothing here has an independent hands-on review yet, and no pricing change was announced with any of the four capabilities. The two betas have no general-availability dates, and Tool Gateway’s beta excludes Azure US. The customer numbers — 7,000 organisations, hundreds of millions of runs a month at Clay, five months to under a day at Workday — are LangChain’s own, not independently measured. The funding is not proof of product quality, and a $3 billion valuation is a bet on the agent-infrastructure market, not a measurement of it.
What is verifiable is the direction of travel. Every announcement this week is about the same shift: from tracing what an agent did to controlling what it is allowed to do, who it acts as, and what it costs per unit of shipped work. That is the difference between a demo agent and a production one, and it is the market LangChain is now explicitly selling into.
Our take: governance is the product now
LangChain’s week reads as a single argument: the hard part of agents was never getting a model to call a tool. It is knowing which agent called which tool, as whom, at what cost, and whether anyone can prove it afterwards. Tool Gateway is the strongest of the four announcements because it answers that question with a mechanism, not a dashboard. Coding-agent ROI tracking is the most useful metric, because cost per merged PR is a number a finance team can actually budget against — provided teams remember that merged PRs are not the only productive work. The trace controls are the most easily misunderstood, because access control is not redaction. If you run agents in production, the sensible first step is small: one agent, one repository, one tool kit, with a deny policy on the dangerous tools and a human approval on the irreversible ones. Then watch whether cost per merged PR moves when you change models. That is the test LangChain’s own week sets up — and the one its customers will now have to pass.
Frequently Asked Questions
What is LangSmith?
LangSmith is LangChain’s agent engineering platform for building, running, improving and governing AI agents. It is framework-agnostic, so it can trace and govern agents built with other frameworks and runtimes, not only LangChain’s own.
How much did LangChain raise in its Series C?
LangChain raised $300 million at a $3 billion valuation, announced on 13 November 2026. The round was led by CapitalG, with participation from GIC, DFJ Growth, IVP, Sapphire Ventures, Salesforce Ventures, Insight Venture Partners, QuantumLight, MongoDB Ventures, Sequoia, Amplify and Benchmark.
Is LangSmith Tool Gateway generally available?
No. Tool Gateway entered public beta on 19 November 2026 on the LangSmith Agent Platform. The beta is available on LangSmith Plus and Enterprise plans, on LangSmith Cloud in the GCP US, EU and APAC regions and AWS US; Azure US is excluded. No general-availability date has been announced.
Does LangSmith for Coding Agents work with Claude Code?
Yes. It supports Claude Code, OpenAI Codex, Cursor, GitHub Copilot in VS Code and Deep Agents. It is in public beta on LangSmith Cloud, and setup is installing the LangSmith plugin for each coding agent, connecting GitHub, and opening the coding-agent overview page.
Do the new trace controls delete sensitive data from traces?
No. They control who can see trace inputs and outputs; the full traces are still stored in LangSmith. Removing sensitive data before upload is a separate SDK feature, using input/output hiding and an anonymiser. Teams handling regulated data need both.
Sources and methodology
- LangChain, “LangChain Raises $300M to Enable Every Company to Own Its Intelligence”, 13 November 2026 (updated 20 November 2026) — https://www.langchain.com/blog/series-c
- LangChain, “LangSmith Tool Gateway: Secure, governed tool use for agents”, published 19 November 2026 — https://www.langchain.com/blog/langsmith-tool-gateway
- LangChain, “LangSmith for Coding Agents: Now track spend by outcome”, published 18 November 2026 — https://www.langchain.com/blog/langsmith-coding-agents-roi
- Unite.AI, “LangChain Launches LangSmith Tool Gateway to Govern Agent Tool Calls”, November 2026 — https://www.unite.ai/langchain-launches-langsmith-tool-gateway-to-govern-agent-tool-calls/
- LangChain on X (@LangChain), 20 November 2026 — https://twitter.com/LangChain/status/2123838785960198512
- OpenAIMaster.ai, “Anthropic’s First Day as a Public Company: ANT Opened at $111.50, Closed at $97.17” — https://openaimaster.ai/anthropic-stock-debut-ant-closes-97-17/
- OpenAIMaster.ai, “How to Use Claude Meetings and Whiteboards: The Rest of Anthropic’s Office Suite” — https://openaimaster.ai/how-to-use-claude-meetings-whiteboards-2026/
- OpenAIMaster.ai, “Cue by Manus: Set Up Your First AI Agent (2026 Guide)” — https://openaimaster.ai/cue-by-manus-setup-guide/
Method: desk research on 21 November 2026. Every factual claim was verified against LangChain’s own blog posts, read in full, with plan and region details cross-checked against Unite.AI and the event recap verified against LangChain’s own X post. Customer figures (7,000 organisations, Clay, Workday) are LangChain’s own and are labelled as such; the $34 billion coding-spend figure is Menlo Ventures’ via LangChain. No hands-on testing was done — LangSmith’s paid tiers were not used for this article — and that is stated plainly. Funding and valuation figures are company announcements, not independent measurements.
Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.


