Back to News An editorial illustration of a glowing browser window carried by a kite over a sea of data, an air traffic map with paused flight paths, and open handwritten notebooks whose symbols rise into a lattice of light
November 15, 2026 Agents Policy Security Funding Research

LangChain Raises $300 Million as Cloudflare Open-Sources a Browser for Agents and Claude Helps Unlock Gödel’s Notebooks

LangChain raised $300 million at a $3 billion valuation to build out LangSmith, and Cloudflare open-sourced Kitesurf, a lightweight browser that lets AI agents load and script web pages on Workers. Officials ruled out the FAA’s new AI planning tool as the cause of Friday’s nationwide ground stop, and a Helsinki logician used Claude to map 663 theorems in Kurt Gödel’s notebooks. Plus: IBM’s injection-resistant Granite previews, the Philippines’ new AI council, Meituan’s LongCat-2.5 weights and Take-Two on AI game clones.

Share

LangChain Raises $300 Million at a $3 Billion Valuation to Expand LangSmith

The biggest funding news of a quiet weekend came from the agent tooling layer rather than a model lab. LangChain’s Series C announcement by Harrison Chase, dated Friday, says the company raised $300 million at a $3 billion valuation. Existing investor CapitalG led the round, with significant participation from new investors GIC and DFJ Growth and existing investors IVP and Sapphire Ventures. The post also welcomes Salesforce Ventures, Insight Venture Partners, QuantumLight and MongoDB Ventures, and thanks Sequoia, Amplify and Benchmark for doubling down.

The money is aimed squarely at LangSmith, which LangChain now describes as a framework-agnostic platform to run, improve and govern agents. The company says more than 7,000 organizations rely on it, naming startups such as Harvey, Rippling and Listen Labs and large enterprises such as NVIDIA, Toyota and Salesforce. Two customer examples carry the pitch: Clay uses LangSmith to monitor hundreds of millions of agent runs per month, and Workday says it cut the time to deploy a new agent from five months to under a day. LangChain splits the product into three pillars: Agent Runtime (Deep Agents, LangGraph, LangChain, Agent Server, Sandboxes, Managed Agents and Fleet), Agent Improvement (observability, evaluation, an in-platform agent called LangSmith Engine and fine-tuning on agent traces) and Agent Governance (LLM and tool gateways for budgets, fallbacks, data protection and tool access). The post does not give revenue figures. LangChain says it will show new LangSmith capabilities at a release event on Friday, November 20.

“To capture that opportunity, every company will need to own its intelligence. 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.”

SEN-X Take

Investors are paying for the neutral control plane, not the original open-source library. If your company runs agents on more than one model provider, the useful question is where traces, evaluations and spending limits live. Keep that layer portable: export your traces, version your evaluation sets, and avoid tying governance rules to a single lab’s console.

Cloudflare Open-Sources Kitesurf, Its Lightweight Browser for AI Agents

Agents spend much of their time reading web pages, and Cloudflare wants that to cost less than running a full copy of Chrome. Cloudflare’s post announcing that Kitesurf is now open source, published Friday, says the stateless browser it introduced in August is now on GitHub under the Apache 2.0 license. Developers can read the code, run it locally, modify it and deploy their own instance to Cloudflare Workers instead of using the company’s managed Browser Run APIs.

“It parses HTML, applies CSS, executes page JavaScript, and renders screenshots without a browser container.”

Kitesurf supports the Chrome DevTools Protocol, so existing clients such as Puppeteer can connect to it, and Cloudflare says agent harnesses can be pointed at the local CDP port. The engine is assembled from open-source Rust packages compiled to WebAssembly: Blitz for HTML parsing, the DOM, layout and painting; Stylo for CSS; Boa for JavaScript, used only as a last resort when scripts cannot run in Dynamic Workers; Linebender libraries for text and 2D graphics; and Krilla for PDF output. The code is organized into three parts, an Engine that manages sessions and client connections, PageScript that runs the page and its scripts, and PageRenderer that draws the result. Self-hosting on Cloudflare requires a Workers Paid plan because Kitesurf creates isolated Workers for pages. Cloudflare cautions that the project is under heavy development and asks contributors to open an issue and be vouched for by a maintainer before sending a pull request.

SEN-X Take

A browser built for agents rather than people could make large-scale page reading much cheaper, but compatibility is the real test. Before moving production scraping or form-filling jobs, run your own target sites through Kitesurf and compare screenshots and extracted data against Chromium. Keep a fallback to a full browser for pages that depend on complex scripts.

FAA Halts Every US Departure for Four Hours, and Officials Rule Out Its New AI Tool

A nationwide ground stop on Friday afternoon briefly turned into an argument about AI in critical infrastructure. Three days earlier, the FAA had expanded its AI-assisted traffic planning system. AeroTime’s report on the FAA expanding SMART to all 22 en route centers says the agency announced the move on November 10, citing 180 rerouted flights in the Washington area with average savings of 10 minutes and nearly 80 gallons of fuel per flight. SMART draws on 200 data streams, and its recommendations pass through FAA specialists and the Air Traffic Control System Command Center before local managers decide whether to act. More than two dozen lawmakers had asked the FAA in October to explain how it tested the system and trained staff, and AeroTime notes the November 10 update did not address those questions.

On Friday the FAA stopped departures nationwide from around 3:54 p.m. until 8:15 p.m. Eastern. AVweb’s account of the ground stop and the SMART questions reports that Transportation Secretary Sean Duffy blamed inconsistent radar data and communications, and officials pointed to overloaded switches and routers in aging telecommunications infrastructure. Newer digital equipment kept working while older analog systems had problems. Homeland Security Secretary Markwayne Mullin said investigators found no evidence of a cyberattack or AI-related interference, though the investigation is ongoing. Duffy said SMART gives recommendations to controllers but is not connected to the STARS or ERAM systems. At least 827 flights had been diverted by early evening, according to Cirium data cited by the Associated Press, and it was the first nationwide ground stop since the January 2023 NOTAM outage.

“There’s no possible way SMART would have impacted what happened,” Duffy said.

SEN-X Take

The evidence so far points to old telecom gear, not the AI tool, yet the timing still forced officials to defend it on live television. Any organization putting AI near critical operations should document its boundaries before something breaks: which systems it can touch, which it cannot, and where its recommendations are logged. That record is what ends speculation quickly.

Claude Helps a Logician Map Hundreds of Theorems in Gödel’s Notebooks

The most striking research story of the week came from the history of mathematics. The Guardian’s report by Alex Bellos on AI and Kurt Gödel’s notebooks describes how Finnish logician Jan von Plato has led a Helsinki University project since 2017 on 16 workbooks Gödel wrote in the early 1940s, 1,009 pages largely in Gabelsberger, an obsolete German shorthand. Progress ran at about a page a week until September, when von Plato gave Claude his transcriptions, six of his own books on Gödel and other secondary literature. Within a couple of days he had a 650-page explanation of the workbooks, which he says had to be checked page by page.

“The AI can connect all the pieces in a thousand pages. It can see all of Gödel’s writings as one piece.”

Von Plato first asked whether the notebooks held a proof of the independence of the axiom of choice, the result Paul Cohen won the Fields medal for in 1966 and that Gödel later claimed he had found in the 1940s. The answer, he says, was almost: Gödel got one definition wrong, though he had the right one two years earlier, and the reconstructed proof passed the Coq proof assistant. Claude then extracted 663 theorems for which Gödel had a proof, a sketch or an idea. Checked in Coq, about 400 were verified as stated, about 200 were repaired with more precise statements or added assumptions, and about 50 were false. Notre Dame logician Joel David Hamkins, who has not seen the reconstruction, said that if it stands up it is remarkable, but that how much is Claude and how much is Gödel is a job for historians.

SEN-X Take

The pattern here travels well beyond mathematics: let a model read an archive no team could finish, then check every claim with a tool that cannot be impressed. Law firms, auditors and research groups sitting on decades of documents should copy the workflow, including the part where roughly one in thirteen extracted results turned out to be false.

IBM Releases Granite Previews Trained to Separate Instructions From Data

Prompt injection remains the most stubborn security problem for agents that read email, documents and tool output. IBM Granite’s Hugging Face post introducing LLMON, published Thursday, describes an LLM-native markup language whose structural delimiters are reserved special tokens rather than ordinary characters such as quotation marks and brackets. IBM released two experimental models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, plus a GitHub toolkit with converters that turn JSON into LLMON and back without loss. Each instruction is explicitly bound to the data it acts on through an execution block, and anything supplied as input, including documents, tool schemas and tool results, is treated as material to read rather than a command.

On the OPI benchmark, which hides injected instructions inside classic NLP tasks, IBM says the LLMON models avoided 47% more attacks than the best frontier model it tested and 95% more than the best open-source model, with similar trends on ASB and StruQ, while general benchmark performance stayed on par with Granite 4.1 Instruct. The models were post-trained from base checkpoints, with decoy instructions added to the data, five GRPO stages and a final DPO stage on a thousand preference pairs built from the model’s own responses to injection attempts. IBM notes the benchmarks test static injections and that it is still working on adaptive attacks that unfold over many turns, and it labels the release an experimental preview not ready for production.

“This is a learned behavior, not an enforced guarantee.”

SEN-X Take

Training a model to know which text is an order and which is just data is the right direction, and IBM deserves credit for stating the limits plainly. Security teams should treat LLMON as an extra layer, not a replacement for least-privilege tool access, human approval on sensitive actions and red-team tests that use multi-turn attacks.

The Philippines Adopts a National AI Governance Framework and Creates an AI Council

Policy news came from Manila. The Philippine News Agency’s report on Executive Order 131 says President Ferdinand R. Marcos Jr. signed the order on November 12, adopting a national AI governance framework and creating the National Artificial Intelligence Council, as Palace Press Officer Claire Castro announced on Saturday. Castro said the measure seeks to give a unified direction for how AI is developed and used, and to ensure it is deployed responsibly, safely and ethically with safeguards that maintain public trust.

GMA News’ breakdown of the order’s provisions adds the structure. The council is chaired by the Executive Secretary, with the Secretary of Economy, Planning and Development as vice chair, and has 10 members: nine heads of agencies including information technology, science, trade, finance, education, labor and budget, plus one private sector representative appointed for a three-year term. It reports directly to the President, must review the framework at least every two years, and will run technical working groups on institutions, data and research, capabilities, infrastructure aligned with the country’s AI Infrastructure Masterplan 2026–2033, and demand generation. Initial funding comes from the existing budgets of member agencies.

“All concerned national government agencies, including GOCCs, are hereby directed, and all LGUs, the private sector, academe, and civil society are encouraged, to render full support and cooperation to the NAIC to ensure the effective implementation of this Order”

SEN-X Take

This is coordination, not regulation: the order directs government agencies and encourages everyone else. For companies with operations or outsourcing teams in the Philippines, nothing changes overnight, but the infrastructure and data working groups will shape incentives, data rules and public procurement. Watch the private sector seat and the first action plans.

Quick Hits: Meituan Opens LongCat-2.5 and Take-Two Calls Out AI Game Clones

Meituan has published the weights for LongCat-2.5. The LongCat-2.5 model card on Hugging Face describes a native multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token, built for long-horizon agent work and trained with optimizations for domestic AI accelerators, under the MIT license. The Decoder’s analysis of the LongCat-2.5 release notes that the self-hosting guide describes a cluster of 194 Huawei Ascend accelerators with no guidance for Nvidia GPUs, and that Meituan’s own table puts the model at 88.4 on Terminal Bench v2.1 against 89.1 for Claude Opus 5.

In games, GamesRadar’s write-up of Strauss Zelnick’s IGN interview reports that the Take-Two chief executive said all companies should protect intellectual property, their own and everyone else’s, and called it ironic for Anthropic and OpenAI to stomp on others’ IP while hoping theirs is protected. GamesRadar notes Anthropic recently added a rule blocking Claude from recreating someone else’s specific game, while OpenAI appears to have no such restriction for now.

SEN-X Take

Both items show capability spreading faster than the rules around it. An MIT-licensed trillion-parameter agent model that runs on Chinese accelerators widens choices for teams outside the US, with export and supply questions attached. And coding agents that can rebuild a game in hours will push studios and AI vendors toward stricter usage policies and more lawsuits.

What We Left Out

Social posts claiming that Anthropic’s IPO order book is about six times oversubscribed, and posts citing OpenAI third-quarter revenue and loss figures, were not confirmed against a filing or a primary report in time, so we did not use them. A reported $300 million Sarvam AI round turned out to describe a June raise, per Sarvam’s Series B announcement, and Mercor’s $750 million Series D post dates to November 10, so neither is new this weekend. A Financial Times report on Nvidia’s chief executive search and Elon Musk posts about Grok changes were not checked against primary sources.

Need help navigating AI for your business?

Our team turns these developments into actionable strategy.

Contact SEN-X →