Nous Research has announced a $90 million funding round and said it will expand its open-source AI agent, Hermes Agent, to businesses. TechCrunch reported on October 7 that the Series B valued the company at $1.5 billion. The effort adds organizational knowledge sharing and cost management to Hermes, which users can run themselves with the model of their choice. The aim is to turn the way a personal agent has learned to work into an asset a company can carry forward. That requires two things: a mechanism for saving procedures that worked, and a management layer that decides who gets to use them.
What "Self-Improvement" Means: Turning Experience into Skills
Hermes Agent's self-improvement centers on saving knowledge gained during tasks to files and reusing it in later work. According to the official skills documentation, the agent can save as skills the methods for multi-step tasks and the fixes that got it past a dead end. When a user corrects its approach, it can also fold that correction into the procedure. A skill here is a set of instructions loaded when the situation calls for it.
Suppose, for example, the agent works out how to produce a weekly report: gathering internal materials, standardizing notation, and finishing in a set format. If that approach is saved, next time there is less need to explain where the materials are or what output format is required. This is an illustration of how the mechanism works, not a measured result from the enterprise version. To understand the growth Hermes promises, it helps to separate this kind of procedure reuse from learning by the model itself.
There is also a separate mechanism for remembering short facts. In the persistent memory documentation, notes about the environment and work, along with user preferences, are saved and loaded at the start of the next session. The saved memory is added to each set of instructions, while longer procedures are read only when needed. The design keeps work experience outside the model.
Saving only helps if it actually happens, though. The same documentation cautions that the agent merely replying that it "remembered" something does not store it; a save operation is required. Memory updated during a conversation is also not immediately reflected in the instructions loaded at the start. Imagining that "self-improvement" means the agent reliably gets smarter with every use would overlook this difference.
Hermes lets users choose the models and execution environment that use the knowledge it has saved. The official overview says it also supports use from messaging apps and scheduled tasks. Know-how is kept around the model rather than fused to a particular one. That portable knowledge is the starting point of the enterprise rollout.
Enterprise Editions: Hand Off Operations or Own the Infrastructure
Official materials describe two delivery formats: Hermes Business, operated by Nous, and Hermes Enterprise, run in a customer's cloud or on-premises facilities. A job posting for a product lead for the business offerings distinguishes between a managed service for teams that want to get started quickly and a form that runs in a customer-controlled environment.
| Format | Where it runs | Positioning in official materials |
|---|---|---|
| Hermes Business | Service operated by Nous | Offered in a form teams can start with easily |
| Hermes Enterprise | Customer's cloud or on-premises facilities | Deployed on infrastructure the customer manages |
This classification, by operator and location, comes from cross-checking the opening of the job posting as of October 8, 2026 against the Enterprise section of the product page. It is not a table showing that all features are available. The choice is between entrusting deployment to Nous and building Hermes into your own infrastructure, and the operational work you take on differs even though the product is the same Hermes.
The product page promotes sharing skills that employees discover across the organization, and visibility into costs across models. The job posting, meanwhile, lists responsibilities for turning those ideas into practical products. Knowledge sharing is described as employees' agents extracting useful procedures and sharing them with the company with permission. Cost management is likewise meant to move beyond showing spending toward judging which model should be used for which task.
The posting also states that the first enterprise deployments are under way. So this is not purely a concept, but it cannot be read as saying that every sharing and management feature has reached general availability. The company appears to be in a phase of rolling out to early customers while building out employee and agent management, auditing, and related functions. Companies considering adoption need to confirm what is actually provided for the features and environment they would use.
Procedures to Share, Secrets Not to Share
To hand a procedure an employee has discovered to another employee, an organization has to manage knowledge sharing separately from access to information. Even if a method for writing a report is useful, not everyone in the company can necessarily read the materials the agent was able to access when it learned that method. Memory that is convenient for individual work becomes a permissions issue inside an organization.
That Nous's job posting says sharing happens "with permission" can be read as a design intent addressing this issue. Questions follow: Does a procedure contain confidential information? How far across departments may it be used? Who corrects a procedure found to be wrong? The more reuse spreads, the more saved content needs to be reviewed and someone needs to be responsible for updating it. This is not a claim that a leak has occurred with this product; it describes conditions for sharing procedures in an organization.
The existing Hermes Agent also does not recommend casually sharing a memory store among multiple agents. The official documentation warns that each agent's writes would get mixed together, and advises separating them into individual configuration units. There are also settings that require human approval for skill changes. A storage mechanism built for individuals cannot simply be turned into common knowledge for an entire company.
Isolating the execution environment must also be considered separately from managing sharing. The security documentation describes user authorization, approval of dangerous operations, and isolated environments. For production use, it calls for settings that limit which users are allowed, cap computing resources, and manage credentials appropriately. Open source means the code can be inspected, but being public does not in itself guarantee safe operation.
Furthermore, running the agent on-premises is a different matter from what information is sent to the AI model you choose. The delivery format of the enterprise edition alone does not support concluding that no data leaves the company. A company must define the range it controls, including where models connect and what permissions the agent has over business tools.
After Adoption, the Question Is Whether the Work Pays
In its funding announcement, Nous gave internal estimates that Hermes has been cloned more than 24 million times and drives roughly 2.5% of the world's AI token usage. Clone counts are not user counts or paying-customer counts, and token usage does not indicate task success rates. For the latter, the relevant section of the announcement does not explain the period, scope, or calculation method. These should be read as the company's own figures on the scale of adoption.
Evaluating enterprise adoption means looking not only at volume used but at the cost of completing work. Even if a cheaper model is chosen, overall spending does not fall if retries increase and people must repeatedly correct the output. What matters is choosing a model suited to the difficulty of the task and measuring whether saved procedures reduce retries. Nous's job posting also lists skill reuse and cost per completed task as evaluation metrics for the enterprise offerings.
NVIDIA, Microsoft's M12, Samsung, and others took part in the round. Nous has said it will use the funds for its enterprise rollout. It can be read as trying to build a business around the publicly released agent, supporting deployment, management, and knowledge sharing. But usage gathered through open development does not necessarily translate into recurring enterprise contracts.
Can a procedure one person learned be used safely by another employee at an adopting company, keep work quality even when the model changes, and lower the cost of getting work done? If a track record builds up on those points, the memory Hermes keeps could grow from a convenient personal note into work knowledge a company can carry forward.