Nous Research Hermes Agent: A Guide to Self-Improving AI

Nous Research is an AI research collective founded in 2023 to build open-source intelligence tools. Hermes Agent, their flagship product released in February 2026, is an autonomous AI agent that learns and improves on its own by converting completed tasks into reusable skills. This guide covers what Nous Research does, how Hermes Agent actually works, and the honest trade-offs versus alternatives like CrewAI and OpenClaw.

Who Is Nous Research?

Nous Research is a team of four co-founders: Jeffrey Quesnelle, Karan Malhotra, Ryan Teknium, and Shivani Mitra. The collective grew out of a Discord community in 2022 and formalized into a company in 2023. They are based in Austin, Texas, and operate with a core mandate: build AI tools that are free, open-source, and accessible to anyone.

Ryan Teknium is well-known in the open-source AI community for earlier model work. The team has built a reputation for shipping fast and iterating based on real user feedback rather than waiting for perfect research papers.

The scale of adoption tells you something. As of October 2026, the Hermes Agent repository on GitHub has 259,500 stars, 57,100 forks, and 60,996 commits. The project went from zero to 95,000 stars in its first seven weeks after launch. That velocity reflects both genuine community demand and a team executing well on a clear problem.

Nous Research operates in the open. All code is MIT licensed. No paywalls. No feature gates. The business model is simple: they offer Nous Portal, a managed subscription service (priced between $29.50 and $59 per month depending on tier), for teams that don't want to run infrastructure themselves. The free version is fully featured and production-ready.

What Is Hermes Agent?

Hermes Agent is a single autonomous AI agent, not a multi-agent framework. It is designed to think, act, and learn independently. At its core, it does three things you should care about:

First, it maintains persistent memory. Unlike a chatbot that forgets your previous requests, Hermes remembers every task you've given it and every solution it has found. This memory is stored in a local ChromaDB instance and remains available for future decisions.

Second, it generates and reuses skills. When Hermes solves a hard problem, it converts the approach into a reusable skill file (in Markdown format) that can be applied to similar problems later. Think of this as the agent teaching itself over time. Each skill is indexed using FTS5 (full-text search), so when Hermes encounters a new task, it automatically searches its skill library first.

Third, it is model-agnostic. You can run Hermes with any LLM: open-source models like Llama via Ollama, proprietary models like Claude or GPT-4, or inference APIs like DeepSeek. The agent framework doesn't care. You pick the model that fits your cost and latency constraints.

Hermes Agent was released as MIT licensed software on February 25, 2026. This means you can use it commercially, modify it, and run it anywhere. You pay for compute (the VPS or server) and the LLM API (if you use one), but not for the software itself.

The Self-Improvement Loop: How Hermes Actually Learns

This is what makes Hermes different from a stateless chatbot. The self-improvement loop is the heart of the product.

Nous Research calls their system GEPA: Goal Expansion, Plan Authoring, and Agent feedback. Here's how it works in practice:

1. You give Hermes a task: "Find three competitors to our SaaS product and extract their pricing pages."
2. Hermes breaks the goal into subtasks, searches its skill library, and executes. If it doesn't find a relevant skill, it creates one from scratch.
3. After completing the task, Hermes reflects on its performance. Did it succeed? What went wrong? What pattern could be reused next time?
4. It then authors a new skill file (a Markdown document) describing the approach, the exact steps taken, and the decision criteria. For the example above, it might create a skill called "web-scraping-competitor-research" with instructions on how to parse pricing tables.
5. On the next similar request, Hermes searches its skill library, finds the skill, and reuses it. Over time, the library grows and the agent gets faster and more accurate.

This is genuinely different from chain-of-thought prompting or few-shot learning in a vanilla LLM. The skills are persistent, searchable, and indexed. You can also audit them, edit them manually, and share them across teams.

The skill library is stored locally, so your proprietary approaches stay in your control. There's no "cloud sync" sending your workflows to Nous Research's servers. This aligns with the open-source philosophy: your data, your rules.

The Five Pillars of Hermes Architecture

Understanding Hermes's design means understanding five architectural pillars:

Memory. Hermes stores task history, completed actions, and learned skills in a ChromaDB vector database. This is semantic memory: the agent can search past experiences by meaning, not just by exact text match. When deciding how to approach a new task, Hermes queries this memory to find similar past solutions.

Skills. Skills are the reusable playbooks Hermes generates. Each skill is a Markdown file documenting: what problem it solves, what inputs it needs, step-by-step execution instructions, and what success looks like. Skills are indexed for fast retrieval and can be combined to solve larger problems.

Soul. This is Nous's term for the personality and instruction layer. You can define how Hermes should think, what values it should prioritize, and what constraints it should follow. This is set once, at the agent level, and guides all downstream decisions.

Crons. Hermes can run scheduled tasks on a timer. If you want the agent to check your email inbox every hour, scrape a website daily, or generate a weekly report, you define that as a cron job. Crons execute independently and feed results back into memory.

Self-Improvement. This is the meta-layer. Hermes can evaluate its own skill library, identify gaps, and request human feedback to refine skills. It can also detect when a skill worked poorly and mark it for revision.

These five pillars work together. Memory feeds self-improvement. Self-improvement generates new skills. Crons keep the system active. Soul ensures the agent behaves as intended.

Hermes Models Explained: 3 vs. 4 vs. 4.3

Nous Research has released several versions of the Hermes base model, each with different trade-offs. Your choice matters for cost and capability.

Hermes 2 was the initial release. It was smaller and less capable than Hermes 3. Most users have moved on; we mention it only for completeness.

Hermes 3 is the stable workhorse. Training statistics: 1 million post-training samples and 1.2 billion tokens. The base model is Llama 2, tuned for instruction-following and reasoning. It comes in multiple sizes: 8B (fits on a single GPU with 16GB VRAM), 70B (requires multiple GPUs or specialized hardware), and in-between variants. If you want to run Hermes locally on modest hardware, Hermes 3 8B is your model.

Hermes 4 is the current flagship. It represents a jump in scale: 5 million post-training samples (5x Hermes 3) and 60 billion tokens (50x Hermes 3). The base model shifted from Llama to Qwen (from Alibaba), which has shown stronger multilingual performance. Hermes 4 was trained on 192 NVIDIA B200 GPUs. It is noticeably smarter but requires more compute: the smallest version is 36B parameters, which needs 24GB+ of VRAM. If you are using a cloud API (Claude, DeepSeek), the underlying model matters less; if you are self-hosting, Hermes 3 8B is more practical for teams without GPU budgets.

Hermes 4.3 (released in December 2025) adds a 36B variant of Hermes 4 and introduces hybrid reasoning modes. Hybrid reasoning means the agent can trade speed for accuracy: on simple tasks, it uses fast inference; on complex ones, it engages a "think deeply" mode with slower but more thorough reasoning.

Here is a comparison table:

Most teams starting out with Hermes Agent use a cloud API (Claude, OpenAI) and focus on the agent framework itself rather than the underlying model. Swapping the model is trivial; it is the persistent memory and skills system that matters.

Licensing, Commercial Use & Trade-Offs

Hermes Agent is MIT licensed. This means:

• You can use it commercially.
• You can modify the source code.
• You can run it on your own infrastructure with zero restrictions.
• There are no compliance certifications on the software itself (Nous Research does not hold SOC 2 or ISO 27001). Your data center provider might; your responsibility is to audit them.

There are honest trade-offs when comparing Hermes to alternatives. Nous Research is not claiming to be the best at everything.

Hermes vs. OpenClaw: OpenClaw is stronger at multi-agent orchestration and planning across teams of agents. Hermes is stronger at autonomous self-improvement and focused execution. Both maintain persistent memory. OpenClaw is better if you need agents to coordinate; Hermes is better if you need a single agent that learns fast.

Hermes vs. CrewAI: CrewAI excels at role-based agent teams (assign roles, define workflows). Hermes is a single-agent framework with built-in self-improvement. CrewAI gives you more granular control over agent roles and responsibilities; Hermes asks less of you upfront and improves through experience.

Hermes vs. LangGraph: LangGraph is a low-level framework for building custom agentic workflows. You have complete control over every step. Hermes is higher-level: you define goals and constraints, and Hermes handles execution and reflection. LangGraph is better if you want precision; Hermes is better if you want autonomy.

Hermes vs. Gumloop or Manus: These are visual, no-code platforms. Hermes is code-first. If your team is not comfortable with Python, Hermes will require learning. The upside: Hermes is more flexible, cheaper to run, and can be integrated into your own systems without a proprietary platform.

The honest take: Hermes is best when you want a single autonomous agent that learns over time and you are willing to manage infrastructure or use a managed service. If you need multi-agent orchestration, role-based workflows, or visual no-code interfaces, another tool might fit better.

Where to Run Hermes: Deployment Options

You have several ways to run Hermes Agent. Each has cost and convenience trade-offs.

VPS with Docker. Rent a $5 to $15 per month VPS from Hetzner, Linode, or DigitalOcean. Install Docker. Run Hermes in a container. This is the cheapest option for production but requires you to manage the server: updates, security, monitoring. If you use an open-source model like Hermes 3 8B via Ollama, your total monthly cost is the VPS fee ($10) plus any API costs ($0). If you connect to Claude API, add ~$8 to $20 per month depending on usage.

Nous Portal (Managed). Nous Research offers Nous Portal, a managed cloud hosting option starting at $29.50 per month (basic tier) and scaling to $59 per month (ultra tier). You upload your agent definition, and Nous handles the infrastructure, updates, and monitoring. This is convenient but more expensive than self-hosting. You are also relying on Nous's uptime, which has been solid but is a dependency.

Local Ollama. If you have a laptop or workstation with 8GB+ VRAM, you can run Hermes 3 8B locally using Ollama (the open-source LLM runtime). This is free and gives you complete privacy (no data leaves your machine). The downside: inference is slow on CPU; you need a GPU for practical use.

Serverless (Modal, Daytona). Modal and Daytona offer serverless GPU runtimes. Deploy Hermes as a function that scales up on demand and tears down when idle. This is useful for bursty workloads (e.g., process uploaded files once a week). Pricing: $0.30 to $0.50 per GPU-hour, which is competitive if you use less than 200 hours per month.

EU Hosting on Opsily. Opsily offers managed Hermes Agent hosting with EU data residency. Your skills library, task history, and memory stay in European data centers. This is important if you have compliance requirements (GDPR data residency, industry regulations). See our pricing breakdown for cost details.

Most teams start with a $5 VPS and Docker to validate the approach. Once you confirm Hermes is adding value, you either scale the VPS or move to a managed service like Nous Portal or Opsily if you need specific compliance or support.

Roadmap & What's Coming

Nous Research has not published a formal product roadmap. However, their release velocity tells you what they prioritize.

Between February and October 2026, they shipped 11 major releases (v0.4 through v0.21.6). That is two releases per month on average. The commit history shows active development across skill generation, memory improvements, and integration with new LLM providers.

Based on GitHub discussions and the features added in recent releases, the team is likely working on:

• Better skill discovery and ranking (finding the right skill faster).
• Community-contributed skills: a marketplace where teams can share and reuse skills.
• Delegation modes: allowing Hermes to hand off subtasks to other agents or humans.
• Plugin architecture: extending Hermes with custom tools and integrations.

But these are inferences, not commitments. The best way to track what is coming is to watch the GitHub releases page and the Nous Research Discord community.

Frequently Asked Questions

What is Hermes Agent?

Hermes Agent is a self-improving autonomous AI agent from Nous Research (released February 25, 2026, MIT license) that runs on your own infrastructure, maintains persistent memory across sessions, and learns by creating reusable skills from experience. It is designed for teams that want an agent to execute tasks independently and improve over time.

How does Hermes Agent learn?

Hermes uses a closed-loop system: it executes tasks, reflects on performance, and extracts reusable patterns into Markdown skill files indexed with FTS5 full-text search. On future similar tasks, Hermes searches its skill library first, reuses winning approaches, and avoids repeating mistakes. Nous Research calls this the GEPA system (Goal Expansion, Plan Authoring, Agent feedback).

Is Hermes Agent free?

The software is free (MIT license) and can be used commercially. Costs come from where you run it (a VPS costs $5 to $15 per month) and which LLM you connect (open-source models via Ollama are free; Claude costs $8 to $20 per month depending on usage). Nous Portal (the managed option) runs $29.50 to $59 per month.

What's the difference between Hermes 3 and Hermes 4?

Hermes 3 has 1 million training samples and 1.2 billion tokens. Hermes 4 has 5 million samples (5x) and 60 billion tokens (50x). Hermes 4 is significantly smarter but requires 24GB+ VRAM to run locally. Hermes 3 8B fits on 16GB, making it practical for cost-conscious teams. Most teams using cloud APIs (Claude, DeepSeek) don't need Hermes 4; the agentic framework matters more than the base model.

How does Hermes Agent compare to OpenClaw?

OpenClaw is stronger on multi-agent orchestration and planning. Hermes is stronger on autonomous self-improvement and focused execution. Both have persistent memory; they have different architectural philosophies. Choose OpenClaw if you need agents coordinating; choose Hermes if you want a single agent that learns fast.

Can I run Hermes Agent locally?

Yes. Docker on a $5 VPS is easiest. Local Ollama requires 8GB+ VRAM for Hermes 3 8B (a laptop GPU works). You can also use Nous Portal (managed cloud) or serverless via Modal or Daytona. The choice depends on your privacy needs, infrastructure budget, and tolerance for self-management.

The Bottom Line

Nous Research is a small, fast-moving team building genuinely useful open-source tools. Hermes Agent is their flagship: an autonomous agent that learns from experience and improves over time. It is free (MIT licensed), runs anywhere, and costs nothing to experiment with.

Hermes is not a fit for everyone. If you need multi-agent orchestration, role-based workflows, or visual interfaces, another tool might be better. But if you want a single agent that improves through persistent memory and reusable skills, and you are comfortable managing infrastructure or paying for managed hosting, Hermes is worth a serious look.

Ready to try it? Start with a $5 VPS and Docker, or explore Opsily's managed Hermes Agent hosting if you need EU data residency and zero vendor lock-in.