Hermes Agent's $90M Series B: Open-Source Runtime Aims for Enterprise AI Control

Funding and Enterprise Pivot

Nous Research raised a $90 million Series B (total funding $158 million) at a $1.5 billion valuation, with participation from Robot Ventures, NVIDIA, Microsoft's M12, and Samsung. The Hermes open-source agent has recorded approximately 22.7 million downloads since February 2026 and over 24 million clones; the company claims Hermes accounts for roughly 2.5% of global token usage, though these figures lack independent verification. CEO Dillon Rolnick announced "Hermes for Businesses," targeting enterprise deployment with a focus on model choice, organizational knowledge retention, and secure, private scaling.

Why Enterprises Need Agent Runtime Control

As AI agents connect to databases, code repositories, and business systems, enterprises invest heavily in tool development, workflow configuration, permission management, and accumulated business experience. If these capabilities are bound to a specific model platform, switching models later could require rebuilding entire workflows. Hermes addresses this by decoupling the execution system from the model, making the model a configurable component while preserving tools, state, and runtime environment independently.

Architecture: Decoupled Components

According to official developer documentation, Hermes comprises multiple interaction entry points (CLI, messaging gateways, API), a shared AI Agent core, a model provider resolution layer, a tool system, and session storage. All entry points invoke the same execution logic, while model inference, tool execution, and state management are handled by separate modules.

Illustrative workflow: A developer requests a code review via Slack. The Messaging Gateway authenticates the user and locates the session; the AI Agent reads context, constructs a prompt, and sends it to the configured model. The model decides whether to call code, file, or terminal tools; the Tool System executes the operations and returns results. The session is saved and the response is delivered back via Slack.

Key architectural separations:

Provider Resolution : Manages model provider, API schema, authentication, and endpoint. Enables swapping models without altering tools or workflows.

Tool System : Independent registration for terminal, filesystem, browser, and external MCP services. New tools can be added without modifying the underlying model.

Execution Backends : Configurable options include local, Docker, SSH, Daytona, and Modal. Enterprises can select isolation levels per workload.

Trade-off: Model replacement still requires re-evaluation of task success rates, tool-call reliability, and context handling. Migration cost is reduced but not eliminated.

Memory and Skills: Persistent Organizational Knowledge

Hermes provides two mechanisms for retaining experience:

Persistent Memory : Stored in MEMORY.md (project conventions, environment) and USER.md (user preferences), with default character limits of 2200 and 1375 respectively. Loaded into the prompt at session start.

Historical Session Retrieval : Full conversations stored in SQLite with FTS5 full-text search. The agent uses session_search to retrieve relevant past discussions on demand.

Skills : Reusable workflows saved via the skill_manage tool. The agent can create, update, or delete skills after completing complex tasks or incorporating feedback. Skills capture executable methods; Memory captures facts and context.

Example: A Data Agent's metric analysis workflow — query specific tables, run validation SQL, check business constraints, generate a formatted report — can be saved as a Skill, while table schemas and business rules reside in Memory.

Governance: Write-approval flags ( skills.write_approval, memory.write_approval) require human review before persisting changes, preventing accidental error propagation. However, file-based knowledge still needs version control, collaboration, permission inheritance — not yet a full governance system.

Limitation: "Continuous learning" relies on external memory and skill updates, not model weight training. More accumulated experience does not guarantee higher task accuracy.

Permission Management: Engineering Guardrails

Hermes uses MCP (Model Context Protocol) to connect external tools (databases, GitHub, internal APIs). Tool configuration limits which operations the model can invoke. Security features include messaging platform allowlists, direct-message pairing, high-risk command approval, file write restrictions, container isolation, and MCP subprocess environment-variable filtering.

Boundaries: File-tool restrictions do not constrain host terminal access. If the agent has shell access, it can bypass file limits. Tool-level permissions alone cannot guarantee safety if the tool itself holds excessive business privileges. Enterprises must align tool permissions, human approval, runtime isolation, and audit logging with actual risk profiles.

Enterprise Value Proposition and Trade-offs

Hermes sells an "open agent runtime" where enterprises:

Choose underlying models

Control deployment environment

Integrate existing business tools

Retain accumulated knowledge (Memory/Skills) in maintainable files

This avoids rebuilding workflows when models change. However, greater control entails greater responsibility: self-managed deployment, monitoring, failure recovery, security auditing, version upgrades, and skill management. Model costs may drop but overall engineering costs may not decrease proportionally.

No public, verifiable enterprise production case studies yet demonstrate task success rates, long-term stability, or TCO advantages. Download/clone metrics do not substitute for production readiness indicators.

Recommended validation: Select a real internal workflow, compare Hermes against the existing solution under identical permissions and tasks on completion rate, failure recovery, model-switch cost, and maintenance effort.

References

1. The Wall Street Journal, "Nous Research Scores $90 Million to Bring Open-Source AI Assistant to Enterprises", October 2026. 2. Nous Research / Dillon Rolnick, "A Note on Our Fundraise", October 2026. 3. TechCrunch / Marina Temkin, "Nous Research confirms it hit $1.5B valuation, launches AI agents for business users", October 7, 2026. 4. Nous Research, Hermes Agent — Official GitHub Repository.