by Chris DePuy, 650 Group / November 19, 2026
On November 18, Nous Research announced a Microsoft partnership for its open-source Hermes Agent. The agent is now offered in three different ways: a desktop client in the Microsoft Store, a chat experience in Microsoft Teams, and containerized execution on Azure (Nous Research press release). The release gives Hermes Enterprise a seat in the Microsoft workplace at each layer, from the desktop to chat to the tenant.

What is Hermes Enterprise?
The commercial offer comes in two forms. Per AlphaSignal launch coverage, Hermes Business is a managed service hosted by Nous Research, while Hermes Enterprise deploys into the customer’s existing Azure environment, operating as hosted agents within the tenant. That is the path for shops with strict rules on data residency, identity or procurement. Which option fits depends on the residency and compliance posture of the organization, not the type of work being delegated.
Hermes Enterprise is the managed, self-improving version of Nous Research’s open-source Hermes Agent, aimed at the enterprise workplace. In the press release, Nous Research states that Hermes Agent tops OpenRouter’s public open-agent rankings, with a reported load approaching 2 trillion tokens per day, as of November 2026. According to the release, employees hand the agent work such as analyzing financial data, drafting documents and clearing a full inbox, then review the finished result in that same chat thread. Agent-generated code executes in Microsoft Azure Container Apps Sandboxes: isolated micro-VMs that boot in well under a second and return compute to the pool when idle.
Dillon Rolnick, CEO of Nous Research, summarized the deal this way: “Hermes became the most popular open agent because it gets better the more you use it. Now teams can use it in the Microsoft tools they already rely on, and what one person figures out becomes shared intelligence the whole company can build on.” Michelle Gonzalez, CVP and Global Head of M12, Microsoft’s Venture Fund, echoed the point, saying the partnership gives organizations a straightforward path to field self-improving agents and grow shared intelligence across the team. Nous Research will also present a session at Microsoft Ignite in San Francisco titled “Run secure agent loops in production with isolated sandboxes at scale.”
The commercial offer comes in two forms. Per AlphaSignal launch coverage, Hermes Business is a managed service hosted by Nous Research, while Hermes Enterprise deploys into the customer’s existing Azure environment, operating as hosted agents within the tenant. That is the path for shops with strict rules on data residency, identity or procurement, and the announcement leaves pricing, SLAs and compliance certifications unspecified. Which option fits depends on the residency and compliance posture of the organization, not the type of work being delegated.
Shared Skills: Collective Wisdom for the Whole Company
The part of the release that most deserves attention is the shared-skills model. Nous Research explains that typical AI tools get one person through a single task, and the judgment that goes into that work, the know-how behind how a company gets work done, rarely travels past the employee who holds it. Hermes Agent works differently. When an employee hands over a recurring job, the agent distills a successful run into a skill — a written description of the approach plus code that executes the stable steps consistently — so a matured workflow draws less inference and runs more reliably with age.
The feature is called Collective Wisdom. A worker can ask the agent which of their skills merit sharing, then put them forward for review and team publication. A teammate who later attempts the same task starts from the strongest approach the team has found, from their first try. Nous Research emphasizes that the resulting knowledge stays a private asset of the company instead of becoming training data for the model vendor. We think that is the durable value in this release: a company’s accumulated know-how becomes a compounding asset that grows with every delegated task, and the skill review flow, which runs like a pull-request approval, is the governance mechanism that lets IT sign off on it. This extends the runtime-security line of work we have covered in Cato’s AgentSense push, since a shared skill library that executes code is a governance surface, not just a productivity feature.
Model Agnosticism: Any Brain, Same Hands
The second notable detail is that Hermes Agent is model-agnostic, which is not the norm for enterprise agents. The Hermes Agent documentation describes configuring inference from cloud APIs, including OpenRouter and Anthropic, down to self-hosted inference on Ollama or vLLM, with routing and fallback between them. The agent framework is MIT licensed, so the loop, the skills and the messaging integrations remain constant while the underlying model changes.
We contrast this with the closed agents that dominate enterprise AI budgets today. An Anthropic-built agent is tied to Anthropic models, and an OpenAI agent to OpenAI models; changing the brain means replacing the agent, its workflows and its integrations, the same lock-in we discussed when Anthropic’s IPO unlocked a new CAPEX wave. With a model-agnostic agent, an enterprise can run the cheapest or most capable model per workload, move inference on-premises where policy requires, and hedge vendor pricing without retraining anyone on a new tool. The model becomes a swappable input; the skills and the workplace integration are what the company owns.
Microsoft Keeps Selling Copilot, Too
The placement itself is the third point. Per the same coverage, Microsoft will field Hermes in its Store and its Teams catalog while keeping its own Copilot lineup on sale, making the agent portfolio in the Microsoft workplace a competitive one rather than a monopoly. For Nous Research, the M12 relationship and a reported $90 million Series B priced at $1.5 billion, per third-party coverage as of November 2026, now translate into one of the largest distribution channels in enterprise software. We have been tracking this same logic across the vendor base, from Cloudflare adding AI workspace partners to HPE extending its self-driving networking push; the platform owner distributes the agents, and the question is who captures the workflow.
How was this calculated?
No 650 Group calculations appear in this post. The dollar figures are the $90 million Series B and $1.5 billion valuation reported in third-party launch coverage, as of November 2026, and the announcement itself leaves pricing, SLAs and compliance certifications unspecified.
We think the enterprise buyer should weigh two things in this release. First, the agent lands in the workplace the organization already has, with tenant-grade control over identity, data and execution. Second, the agent’s durable value, shared skills and model choice, accrues to the company rather than to a single model vendor. The coming deployments will show whether an open, model-agnostic agent can win the managed-workplace default. 650 Group researches several of the technologies we covered in this blog, including Enterprise and Security and SASE.