An open-source agent that started as a developer tool now has a front door in the Microsoft workplace. On November 18, 2026, Nous Research announced in a press release, Nous Research partners with Microsoft to bring Hermes Agent to the enterprise, that Hermes Agent works in Microsoft Teams, in the Microsoft Store, and on Azure Container Apps Sandboxes. The news landed during Microsoft Ignite, where Nous presented in session BRK448.
The facts come from Nous itself, so treat the usage numbers as company claims. Below are the product pieces, the data details from Nous's own Teams help page, and the questions the announcement leaves open.
TL;DR: what is announced and what is not
| Item | What Nous says |
|---|---|
| Hermes Desktop | Available for Windows in the Microsoft Store |
| Teams | Employees hand Hermes work, such as financial analysis, drafting documents or inbox triage, and review the result in the same conversation |
| Azure | Agent workloads run in Azure Container Apps Sandboxes: isolated microVMs that start in under a second and release compute when idle |
| Collective Wisdom | Employees propose and publish skills to their team, with review |
| Hermes Enterprise | Can run in the customer's own Azure tenant under a similar product and pricing model |
| Scale claim | Number one among public open agents on OpenRouter, nearly 2 trillion tokens a day |
| Not published | Price, tenant deployment steps, certifications, a customer reference |
What Hermes Agent is
Nous describes Hermes Agent as an open-source, self-improving agent "used by millions of people and organizations worldwide." The core idea, per the release: as an employee delegates recurring work, Hermes turns a successful approach into a reusable skill, "a written definition paired with code that runs the stable steps the same way every time." The claimed benefit is that established workflows use less inference and become more reliable over time.
We have covered Hermes from several sides. See the harness reliability leaderboard work on Hermes Agent, the Hermes mobile beta, claimed versus verified, and the OpenRouter token run-rate claim that gives context for the "number one app" ranking. The "nearly 2 trillion tokens a day" figure is Nous's own statement; we could not verify it independently.
The three Microsoft pieces
Hermes Desktop in the Microsoft Store
Hermes Desktop for Windows is listed in the Microsoft Store. The practical gain for IT is distribution: a Store app can go through the same channels as other managed software. Nous's press release does not describe a managed-deployment guide, so check the Store listing and your admin tooling.
Microsoft Teams
In Teams, employees assign knowledge work to Hermes and read the finished result in the same chat. The Nous Portal's help page gives the details a Teams admin needs:
- Permissions. Hermes Agent is a bot. It "does not ask for Microsoft Graph or resource-specific consent permissions, and it does not sign users in to Microsoft."
- What it sees. It receives only messages sent to it in a 1:1 chat and messages that @mention it in group chats and channels.
- Admin control. Admins allow or block it in the Teams admin center under Manage apps, and use app permission policies and setup policies to limit who gets it.
- Linking. Each user connects their own Teams account to their own Nous account with a link code that expires in 10 minutes. Codes are not shared.
- Limits. Files work only in 1:1 chats. Guests and people from other organizations are not supported yet.
Azure Container Apps Sandboxes
Agent code runs in isolated microVMs that start in under a second and release compute when idle, per Nous. For security teams this is the relevant part: agents that execute code need isolation. We compared other options in our agent sandbox comparison, and Microsoft itself has been publishing agent tooling around Ignite; see the complete Ignite 2026 coverage.
Collective Wisdom: shared skills with review
Collective Wisdom is the feature that distinguishes the enterprise pitch from single-user Hermes. Employees can ask Hermes which of their skills are worth sharing, then propose and publish them to their team with review. Nous's framing: "The next person who attempts the task starts from the best approach anyone on the team has found, on day one." It also says the resulting shared intelligence "remains a private asset of the organization rather than training data for an AI vendor."
That last sentence is a data-use promise. It is worth turning into a contract term. Ask Nous to confirm it in writing for your tenant, along with who can see skills before they are published and how a review step is enforced.
Where the data goes: two different stories
The press release and the help page describe two deployments, and they differ.
- The Teams app through Hermes Cloud. The help page says the agent "runs on Hermes Cloud in the Nous Portal." Messages and files from 1:1 chats pass through the Nous connector service to your agent, along with your Teams account and organization IDs and display name. The connector keeps a reference to each Teams conversation so it can deliver replies, and keeps the link between your Teams and Nous accounts until you disconnect. Conversations stay in your agent until you delete it.
- Hermes Enterprise in your own Azure tenant. The press release says organizations with strict data residency, security or procurement needs can run it in their own tenant. This is the option for regulated companies.
If you evaluate the Teams app, read the help page's data section as the baseline. Whether the own-tenant option removes the Nous connector from the message path is not stated in the sources we read. Ask. For a broader pattern of how a large company governs agents across teams, compare BMW's approach, which keeps governance in a gateway and registry outside the agent.
The funding and partner context
The release describes Nous as an M12 portfolio company, M12 being Microsoft's venture fund, and quotes M12's Michelle Gonzalez and Nous CEO Dillon Rolnick. That relationship is disclosed in the release. It matters for how to read the announcement: it is a partnership with a portfolio company, announced at Microsoft's own conference, not an independent Microsoft product launch. Microsoft's own agent efforts are separate; for those see our post on Microsoft's open-source Amplifier adapters for Slack, Teams and other agents.
Nous also publishes research on its blog, including a post-trained Qwen3.8-27B model reinforcing Western values, which we checked against its own numbers. Readers who evaluate Nous as a vendor should weigh both.
A two-week pilot plan
A press release cannot tell you whether an agent helps your team. A small, measured pilot can. Here is a plan that uses only features Nous describes.
- Pick three recurring tasks. The release names financial analysis, drafting documents and inbox triage. Choose tasks with a clear right answer you can check, such as a weekly report with fixed fields.
- Define success before you start. For each task, write down time saved, error rate against a human version, and how often a person had to redo the work. Without a baseline, "it felt faster" is all you will have.
- Run two people on each task. Let the first person delegate to Hermes for several rounds. Then let the second person use the skill Hermes created, through Collective Wisdom, and compare their results. This tests the central claim that the next person starts from the best approach.
- Review every published skill. The review step is where a bad habit could spread to a whole team. Read the written definition and the code before approval.
- Check the data path. Send a test message that contains a harmless marker string, then ask Nous where it is stored and how long. Compare the answer with the help page.
- Decide on the tenant option. If the pilot works and your compliance team requires residency, ask for the own-tenant terms and pricing in writing.
Keep the pilot small, and keep a person in the loop for anything that leaves the team, such as an email draft sent to a customer.
What is not published
- Price. "Similar product and pricing model" is the only statement. There is no figure.
- Tenant deployment details. Which Azure services, what the connector does in your tenant, and how updates work are not in the release.
- Compliance. No certification, audit or data-processing terms are cited.
- Customers. No named enterprise customer or measured result appears in the release.
- Safety controls for Teams actions. The help page notes responses "may be incorrect." What the agent can do beyond chat, such as run code on your behalf, depends on your configuration.
What to do now
IT and Teams admins: Read the help page, then pilot the bot with a small policy group. Confirm it only sees messages that mention it, and decide whether to allow files in 1:1 chats.
Security teams: Ask Nous to document the data path for both deployments, the retention of the connector's conversation references, and who can approve a skill in Collective Wisdom.
Founders and team leads: The shared-skills idea is the part to test. Pick one recurring task, have two people run it, and see whether the published skill produces the same result for the next person.
Sources and related reading
- Nous Research, Nous Research partners with Microsoft to bring Hermes Agent to the enterprise (November 18, 2026)
- Nous Portal, Hermes Agent for Microsoft Teams help page
- Hermes Agent harness reliability
- Microsoft Ignite 2026 coverage
- Agent sandboxes compared
