Who is building the AI agent stack? The market in November 2026
LangChain at $3 billion, Modal at $15.75 billion, MCP made stateless and 305 members in its foundation. Where the money and the standards went in 2026.
In 2026 investors paid most for the layers underneath AI agents. They backed the places agents run code, the frameworks that run them, the gateways that control what they reach, and the tools that check what they did. LangChain raised $300 million at a $3 billion valuation on 13 November. Modal, which runs code in sandboxes for agents, raised $750 million at $15.75 billion in October. Large security companies bought the startups that manage agent identity. The standards moved into one place, the Linux Foundation's Agentic AI Foundation, which now has 305 members. And a survey by McKinsey found that 40% of large companies are now scaling agents in at least one part of the business.
What did LangChain raise, and what is it for?
A $300 million Series C at a $3 billion valuation, led by its existing investor CapitalG. TechCrunch reports that new backers include GIC, DFJ Growth, Salesforce Ventures, Insight and MongoDB Ventures. Its previous round, about a year earlier, was $125 million at $1.25 billion.
LangChain did not disclose revenue. It now presents its commercial product, LangSmith, in three parts: a runtime that runs agents, tools to observe and evaluate them, and governance through gateways for models and tools. It says more than 7,000 organizations use LangSmith. The third part is the newest, and the most telling. A framework company now sells control over what agents can reach.
Why are agent sandboxes worth so much?
Because every coding agent needs somewhere to run code that is not somebody's laptop, and that is now a large business.
Modal's Series D on 14 October was $750 million, led by Accel, at a $15.75 billion valuation. Modal says it passed $1 billion in annualised revenue, with sandboxes alone at $300 million, and that about 3 billion sandboxes have been created on it. Daytona, a smaller rival, raised $60 million at a $1 billion valuation on 20 October.
A sandbox solves where the code runs. It does not decide what the code may reach, which is why the next layer is attracting buyers.
Who is buying agent security and identity companies?
The large security and data companies, mostly by acquisition.
- Palo Alto Networks completed its purchase of CyberArk, an identity security company, on 11 February 2026. It framed machine and agent identity as a core reason.
- Snowflake agreed in May to buy Natoma, a startup that controls which systems agents can reach over MCP.
- ClickHouse bought Langfuse, an open source tool for tracing and evaluating AI applications, in January. Langfuse stays open source.
The common thread is that agents are a new kind of identity. Each one holds access, acts without a person typing, and needs its own record of what it did.
What is the Agentic AI Foundation?
A Linux Foundation body that now holds most of the open standards agents use. It was formed on 9 December 2025 around three donated projects: MCP from Anthropic, goose from Block and AGENTS.md from OpenAI. Its platinum members are AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.
It grew fast in 2026. Google's agent-to-agent protocol, A2A, joined in August. On 22 October the foundation added four more projects: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. The same day it said membership had passed 300 organizations, at 305.
Rival labs putting their standards in one neutral home means a team can build on them without betting on one vendor.
What changed in MCP this year?
The protocol became stateless. The specification released on 28 July 2026 removed the opening handshake and the session header, so any server can answer any request. The main changes:
- New headers that name the method and the tool on each request, so a gateway can route and check a call without reading its body.
- Cacheable lists, so a client can keep a server's list of tools for a stated time.
- Long-running tasks and interactive apps moved into official extensions.
- Stricter sign-in, with checks on which server issued a token.
- Older features deprecated, including sampling and roots, with at least 12 months of support.
Adoption kept growing. By September, downloads of MCP's main software kits were approaching half a billion a month. The official MCP registry is still in preview. We compare MCP with A2A in A2A vs MCP: what each agent protocol is for.
How many companies are actually using agents?
More than last year, and big companies far more than small ones. McKinsey's 2026 State of AI survey, of 1,719 people from May to June, found that 40% of organizations with more than $1 billion in revenue are scaling agents in at least one function, up from 27%. Among smaller organizations the share stayed at 22%. About one in five respondents is scaling coding agents.
The money is harder to find. About 37% said AI had added to their company's operating profit, no change from 2025. About 20% said running costs, including tokens, held back their use.
Gartner's warning from June 2025 still frames the debate. It predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of cost, unclear value or weak risk controls.
What does this mean for a team choosing agent tools?
The stack has settled into recognisable layers, and each one solves a different problem.
- Where the agent runs. A sandbox or microVM, so its code cannot touch your machines.
- What the agent can reach. A gateway that holds credentials, allows tools explicitly and records every call. This is where the 2026 acquisitions were.
- How the agent performs. Tracing and evaluation, to improve prompts and catch regressions.
- Whether the work counts. A check of each piece of work against what was asked, before it is accepted.
Choose tools that speak the open standards, MCP above all, so that a layer can be replaced without rebuilding the rest. Gartner's reasons for cancellation are unclear value and weak controls, and those are the last two layers.
Cognibl covers the reach and the result. Agents connect to your tools through its MCP gateway, which holds every credential, refuses anything not explicitly allowed and records each call. A task only moves to done with a proof of work that is checked against its definition of done. See where Cognibl sits in the stack.
Keep reading
- What is OpenAI worth, and when will it go public?OpenAI raised money at a $1.25 trillion valuation in November 2026 and says it will list in 2027. Its owners, revenue, and how it compares with Anthropic.
- Should you run coding agents on open-weight models?Open-weight models cost less per token than Claude or GPT, but not always per task. What their licences allow, and what tests found on agent hijacking.
- Jensen Huang has left Nvidia. What happened, and what it means for AIJensen Huang resigned as Nvidia's CEO on 12 November 2026. What the filing says, the record quarter that followed, and the new US rule on AI chip access.