The most important detail in Nous Research’s new funding round is not the money. It is what investors are backing: an AI agent that can work across files, browsers, code environments and language models. That is a different product category from computer vision or conventional natural language processing. It moves AI from generating an answer to carrying out a sequence of actions.
Nous Research, developer of the Hermes agent, announced a $90 million Series B led by Robot Ventures, with participation from Nvidia, Samsung, Y Combinator and others. The company is reportedly valued at $1.5 billion, while Hermes has been downloaded more than 24 million times and accounts for an estimated 2.5% of global token consumption. These figures are reported by SiliconANGLE and should be treated as indicative, subject to the source’s methodology and later revision.
For founders, the practical question is not whether to add an agent feature. It is whether the product can safely coordinate real work across tools, data and models.
The product boundary is moving from “generate content” to “complete a workflow.”
Why computer vision and natural language processing are not the main shift
Computer vision helps software interpret images. Natural language processing helps software interpret and generate language. Both remain important capabilities, but Hermes illustrates a higher-level layer: an agent can decide which model to use, access a browser, run code inside a Docker container, interact with local files and preserve selected context across sessions.
That distinction matters because the user is no longer buying an isolated model response. They are delegating a bounded process. For example, “organize these documents” may involve inspecting file types, classifying content, creating folders, renaming files and reporting exceptions. Each action creates a new reliability, permission and audit requirement.
Nous appears to address this through a model router that considers cost, speed and output quality. It also offers personalization through preferences and optional external databases, including the agent-focused RetainDB store. The advantage is flexibility. The trade-off is a larger failure surface: more models, data stores and tools mean more points where permissions, context or outputs can go wrong.
The hidden signal in the Nous Research funding round
What caught my attention is the combination of open distribution and enterprise controls. Hermes was released under an open-source license, and the company says it has attracted several thousand contributors. At the same time, its paid business editions include administrative controls such as team-level token limits and shared agent customizations.
That combination suggests a familiar commercial pattern. Open distribution can accelerate experimentation and developer adoption. Enterprise buyers, however, generally pay for governance: identity management, usage policies, data boundaries, observability, support and predictable deployment. Open source can widen the top of the funnel, but it does not remove the cost of operating an agent in a business environment.
The company’s Hermes 4 models were customized using 5 million synthetically generated records, with about two-thirds created to improve reasoning, according to the source. That points to a second strategic lesson: model differentiation may come less from training a larger general model and more from building a targeted data-generation and evaluation system.
Figures mentioned are indicative and may vary depending on reporting methodology, definitions, source verification and market conditions.
What founders should learn from the shift
Historically, the internet became commercially powerful when it moved beyond publishing pages and started coordinating transactions, identity and logistics. Cloud software followed a similar path: the important product was not simply remote computing, but a layer that made complex resources usable through repeatable workflows.
AI agents are following that pattern. Models provide reasoning and generation, but the durable product may be the control layer around them. That layer decides what the agent can access, which model handles a task, what gets stored, when a human must approve an action and how the result is evaluated.
The effect chain is straightforward:
- Agents connect language models to business tools and private data.
- Workflows become easier to delegate, but errors can propagate across multiple systems.
- Buyers shift attention from model quality alone to permissions, traceability and operational controls.
- New businesses emerge around deployment, evaluation, data boundaries and workflow-specific reliability.
This is why the round matters beyond Nous Research. Nvidia and Samsung are not only participating in another model company’s financing. Their involvement is consistent with a market where demand may increasingly sit at the intersection of models, devices, compute and software that manages execution.
A practical framework for deciding whether to build an agent
Do not start with “Where can we add AI?” Start with a workflow that has a clear owner, repeatable inputs and an observable definition of success. Then assess it against four questions:
- Permission: What systems, files or accounts must the agent access, and can access be limited to the minimum required scope?
- Failure: What happens when the agent misunderstands a request, receives incomplete data or chooses the wrong tool?
- Evaluation: Can the team test outputs against known examples before the workflow reaches a customer or colleague?
- Economics: Does routing between models reduce total cost enough to justify the engineering, monitoring and support burden?
A document-classification agent may be a reasonable first use case because inputs and outputs can be sampled. An agent that sends customer messages, changes financial records or modifies production systems requires stronger approvals and rollback paths. The benefit of automation must be weighed against integration work, data handling obligations and the cost of investigating mistakes.
Where the supporting layer is still underbuilt
The immediate opportunity is not another general-purpose chatbot. It is software that makes agents accountable inside a specific operating environment. Examples include approval queues, tool-permission gateways, evaluation suites, cost routers, context stores and workflow logs.
Teams working on prompt engineering should connect prompts to test cases rather than treating them as permanent instructions. Teams exploring reinforcement learning should define what feedback is reliable and who is responsible for reviewing it. Security teams should treat agent tool access as a new surface for penetration testing, not as an ordinary software integration.
What does "ci cd pipeline" mean for agent products?
If you typed "ci cd pipeline" into a search engine, you are probably looking for a way to test and release software consistently. For agent products, that means versioning prompts, model choices, tools, permissions and evaluation datasets alongside application code. An agent change should be reviewed for behavior and access impact, not only whether the build passes.
This is also where ai agents become an operational discipline rather than a demo category. The team needs clear boundaries for what the system may do, what it must ask permission to do and what it must never do without human confirmation.
What to do before adding an agent to your product
Choose one workflow and document its current steps before selecting a model. Record the tools it touches, the data it needs, the acceptable error rate for the business context and the point at which a person takes over.
Then run a small evaluation using representative cases, including ambiguous requests and deliberate edge cases. Compare a single-model design with a routed design only after measuring latency, cost, output quality and debugging complexity. Routing may improve economics in some workloads, but it can also make behavior harder to reproduce.
Finally, design the commercial boundary early. A consumer plan may prioritize ease of setup. A business plan may need usage limits, shared configurations and administrator visibility. An enterprise deployment may additionally require private networking, retention controls, audit records and integration support. These are different products, not simply higher pricing tiers.
Nous Research is using its funding to expand adoption of Hermes’ business editions, according to the source. Whether that strategy works will depend on converting broad experimentation into repeatable, governable workflows. The funding round is therefore best read as a signal about where value is moving: away from model access alone and toward systems that can act inside a company’s real operating environment.
For more practical analysis on AI, software and execution, explore the Yanisa Execution blog or compare your current automation priorities with the framework above.
Frequently Asked Questions
What is Hermes by Nous Research?
Hermes is an AI agent developed by Nous Research. It can work with local files, browsers, code environments and multiple language models, depending on the configuration and permissions provided.
Why did Nvidia and Samsung invest in Nous Research?
The source reports that Nvidia and Samsung joined Nous Research’s funding round alongside Robot Ventures, Y Combinator and others. Their participation may reflect growing interest in software that connects AI models with devices, compute and business workflows, but the investors’ specific rationale was not detailed in the source.
How does an AI agent differ from a chatbot?
A chatbot generally responds to a user’s message, while an agent can be configured to plan and execute multiple steps using tools. The agent still needs defined permissions, testing and human oversight appropriate to the task.
What does model routing do in Hermes?
Model routing selects among available language models for a prompt. The source says Hermes considers factors such as inference cost, speed and output quality when making that selection.
Is Hermes available for businesses?
The source describes Hermes Business and Hermes Enterprise editions for organizations. These editions include controls such as team-level token limits and shared agent customizations, with the exact capabilities depending on the plan and deployment.