- Towards AGI
- Posts
- AI Agents Just Became A $3B Business. Who Governs Yours?
AI Agents Just Became A $3B Business. Who Governs Yours?
7,000 companies now run agents on one platform. here's what you should prepare for
Today, we're diving into:
Gen AI: LangChain raised $300M at a $3B valuation, and more than 7,000 organizations now run agents on its LangSmith platform
Hot Tea: Mirendil, started by former Anthropic researchers, raised a $1B Series A to build AI that does AI research
OpenAI: OpenAI has disbanded its stand-alone robotics team as its general-purpose models start running robots
AI Agents Just Became A $3B Business. Who Governs Yours?
LangChain has raised $300 million at a $3 billion valuation, CEO Harrison Chase announced on November 13. CapitalG led the round, with GIC, DFJ Growth, IVP and Sapphire Ventures joining, TechCrunch reported.
A year ago, the same company was valued at $1.25 billion. What changed is not the hype around agents. It is how many companies now run them in production.
Your Agents Are Only As Good As What They Can Trust.
Find out whether your data is ready for agents that act on it every day.

What The Money Is Buying
LangChain started in 2022 as an open-source project. Today its focus is LangSmith, software for building and managing agents that the company says is not locked to any model, cloud or ecosystem.
The new money goes into three pillars: Agent Runtime, Agent Improvement and Agent Governance. The company says more launches are coming on November 20.
"Every company will need to own its intelligence." Harrison Chase, CEO, LangChain
The customer list explains the valuation. TechCrunch counts 7,000 LangSmith customers, including Harvey, Rippling, Clay, Nvidia, Toyota and Salesforce.
The Numbers Behind The Shift
Clay monitors hundreds of millions of agent runs a month on the platform.
Workday cut the time to deploy a new agent from five months to under a day, according to LangChain.
When building an agent takes a day instead of five months, the hard part moves. It is no longer whether your teams can build agents. It is whether anyone can keep track of what all of them are doing.
This Is Not A Framework Problem. It Is A Governance Problem.
The easy reaction is to pick a winning agent platform and standardize on it. That helps, but it misses where agents actually fail.
An agent built in a day still reads your data, your definitions and your permissions. If those are scattered, stale or owned by nobody, a faster build only means faster mistakes.

Notice that governance is now one of LangChain's three pillars. The vendors have worked out that control, not creation, is the next bottleneck.
The real bottleneck is not the number of agents you can launch. It is whether each one reads from context the business has agreed is true, and whether you can see what it touched.
The better alternative is to treat agents like a new class of employee. Give each one a defined role, a list of approved data and a trail you can audit.
What You Should Do This Quarter
Count your agents. Include prototypes and team-built tools, not just the ones IT approved.
Keep your platform swappable. Choose tooling that is not tied to one model, so you can change models as prices and quality move.
Name a data owner per agent. Someone should be accountable for every source an agent reads.
Monitor runs, not just launches. Agents fail quietly in production, long after the demo went well.
Building agents you can govern starts with trusted enterprise context. DataManagement.AI helps organizations create that foundation by unifying metadata, lineage, governance, business glossaries and knowledge assets into a single intelligence layer. It automatically discovers metadata, maps lineage across the modern data stack and identifies downstream dependencies, creating a centralized knowledge layer that stays continuously updated as your environment evolves.
The organizations moving first will not just build agents faster. They will know what every agent reads, who owns it and what it did last night.
Ex-Anthropic Researchers Raised $1B To Automate AI Research. Is Your Roadmap Ready?
Mirendil has raised $1 billion in a Series A led by Kleiner Perkins and Clearlake Capital, with Google among the participants, the company announced on November 12.
The company is less than a year old. Its goal is to build AI that plans, runs and learns from AI experiments with far less human help.
Who Is Behind It
CEO Behnam Neyshabur co-led Google DeepMind's Blueshift team, and CTO Harsh Mehta built the first version of Anthropic's internal AI research platform, according to Dealroom. Both left Anthropic in December 2025 and founded Mirendil that month.
"The most important application of AI is AI itself." Behnam Neyshabur, CEO, Mirendil, on the company's seed round
In June, the company raised a $200 million seed round at a roughly $1 billion valuation, SiliconANGLE reported. Five months later, it has five times that amount.

What Mirendil Says It Has Built
Mirendil-Base: a coding model the company says it pretrained from scratch in 3.5 months, using under $10 million of compute.
Fern: a post-trained version with up to 10 million tokens of context.
Lor: a private-beta product in which hundreds of agents write code and run experiments for days.
These are the company's own claims, and none has been independently tested yet. But the direction is what matters for your planning.
This Fundamentally Changes The Speed Of Your Vendors
If agents can run AI experiments around the clock, the cycle between model releases gets shorter. That is good news for capability, and harder news for anyone who signed a three-year AI roadmap built around today's models.
The wrong reaction is to wait for the frontier to settle before committing. It will not settle. Every quarter you wait, the models you are evaluating are already a generation old.
The better alternative is to design for change. Build your AI stack so models can be swapped, tested and retired without rebuilding the applications around them.
Separate apps from models. Put a gateway or abstraction layer between your workflows and the model behind them.
Keep your own test set. A fixed set of your real cases lets you judge each new model in days, not months.
Watch the cost of training, not just inference. If a capable base model can really be built for under $10 million, owning a model becomes a realistic option for more enterprises.
The organizations moving first will not just follow the frontier. They will build a stack that can absorb a better model the week it arrives.
OpenAI Just Disbanded Its Robotics Team. Its General Models Took Over.
OpenAI has broken up its stand-alone robotics team, Fortune reported in an exclusive on November 13. Robotics researchers now sit in the main research team under Chief Research Officer Mark Chen. Applied robotics and hardware staff move to the Applied organization under President Greg Brockman.
OpenAI told Fortune that robotics remains a core part of its long-term vision. Fortune also reported that Sam Altman is targeting a first public demo of an OpenAI humanoid in 2027.
Why Fold The Team Now?
Fortune's headline gives the reason: OpenAI's general-purpose models are starting to run robots themselves. A separate team building separate robot brains matters less when the main model can do the job.
Fortune cited an Nvidia study in which an Astra-powered agent reached a 74.8% success rate on a simulated manipulation benchmark, against 51.0% for the best specialized robotics models on the leaderboard.
Carnegie Mellon's Guanya Shi told Fortune he would read the move as robotics being "no longer a special problem."

The Pattern Is Bigger Than Robots
This is the same story enterprises keep seeing in miniature. A specialized model is built for one job. A year later, a general model matches or beats it, and the specialized one becomes an expensive thing to maintain.
That does not mean specialized models are dead. It means the case for each one has to be proven on your own work, and proven again every time a new general model ships.
The Real Bottleneck Is Measurement
The wrong reaction is to swing from one extreme to the other: scrap every custom model, or keep funding all of them because they once won a pilot.
The better alternative is to measure. Run the general model and the specialized one side by side on the same tasks, and compare accuracy, speed and cost per completed job, not per token.
Know Your Inference (KYI) is built for exactly this decision. When you measure how accurate, fast and costly each AI workload is on your own cases, you can see which jobs still earn a specialized model and which a general one now handles better and cheaper.
List your specialized models. Note what each one does and what it costs to keep running.
Re-test after every major release. Put each new general model against your specialized ones on the same cases.
Retire with evidence. Move a workload only when the numbers show the general model wins on your work.
The organizations moving first will not just bet on general or specialized AI. They will measure both on their own work and move the moment the numbers change.
Journey Towards AGI
Research and advisory firm guiding on the journey to Artificial General Intelligence
Know Your Inference Maximising GenAI impact on performance and Efficiency. | Model Context Protocol Connect with us, and get end-to-end guidance on AI implementation. |
Your opinion matters!
Hope you loved reading our piece of newsletter as much as we had fun writing it.
Share your experience and feedback with us below 'cause we take your critique very critically.
How's your experience? |
Thank you for reading
-Shen & Towards AGI team
