Video summary
Nous Research’s Hermes Agent: The Case for Open Models in Production | Arize Observe 2026
Main summary
Key takeaways
Summary
Sam Herring of Nous Research presents Hermes Agent, an open-source, command-line-first AI agent framework. It began as an internal productivity tool after the team sought a different approach to existing agent software, particularly for longer-running work and persistent memory. Nous later released it publicly and made it a central product.
Main features and design
- Local and provider-flexible: Hermes runs on a user’s computer, connects to tools, and works with model providers including Anthropic, OpenAI, OpenRouter, and Nous’s API. It can also use locally hosted models through options such as vLLM, LM Studio, and llama.cpp. Herring says local models can support offline or air-gapped use while keeping data private.
- Tools and skills: The framework includes more than 80 tools for common operations, such as running code and reading files, plus roughly 80 bundled skills. Skills are Markdown descriptions of repeatable workflows that can be shared or adapted across agent frameworks.
- Automatic skill creation: Hermes can recognize when a workflow may be useful again, save it as a skill without being explicitly asked, and update it when the workflow changes.
- Multiple memory levels: Herring describes short-term, session-related memory; conversation-level files; and user-level, cross-session memory backed by a vector database. The system is designed to preserve information according to its likely future usefulness. He also mentions seven memory providers.
- Long-running and asynchronous work: Messaging integrations include Telegram, WhatsApp, Signal, and iMessage. Users can send tasks to the agent and receive results later. Scheduled tasks can run in the background and send notifications.
- Agent loop and efficiency: The core agent loop is described as about 12,000 lines of Python. It follows a typical cycle: a user message goes to an LLM, the model replies or calls tools, tool results return to the model, and the agent produces a response. Hermes also supports manual and automatic conversation compression, along with query caching intended to make repeated skill workflows more efficient.
- Extensible and community-driven: Users can ask Hermes to inspect its codebase and help implement missing features, then submit the work as a pull request. Herring encourages customization and community contributions.
Examples and use cases
Herring shares several examples from his own use and the Hermes community:
- Scheduled car search: Hermes searched Facebook Marketplace, auction sites, and other listings for a specific vehicle. It then ran the search each morning and notified Herring when it found a match. He says this eventually helped him buy a car.
- Visual clothing search: Herring built a pipeline that embedded images of clothing listings and compared them with a reference image of pottery. He could send an image through Telegram and receive visually similar listings with similarity scores, plus filters such as size and price.
- Research and coding: Herring says he uses Hermes for production coding and research. He describes sending experiment requests asynchronously and receiving results, including ablation results, while away from his computer.
- Reinforcement learning experiments: He connected Hermes with a reinforcement-learning setup to create an environment, train a model, and reload the resulting weights into an agent instance. He presents this as an early exploration of ongoing learning or recursive self-improvement, not as a finished capability.
- Community projects: Examples include creative work such as music and video generation, smart-home automation, research assistants, and hackathon projects. One smart-home project tracked how often a dog door opened.
Guides, reviews, and practical information
- Installation guidance: Herring points viewers to the project’s GitHub repository and documentation for installation instructions, with options discussed for Mac, Windows, and Android.
- Trials and model support: He mentions periodic free model trials and says Nvidia’s Nemotron 3 Ultra had recently been tested with Hermes and performed well in internal benchmarks.
- Review evidence: The talk is a product presentation rather than an independent review. Herring offers personal endorsements and examples of his own use, but does not present a formal comparative benchmark or independent evaluation.
Main speaker and source
Sam Herring, machine learning engineer at Nous Research, is the presenter and source of the product descriptions, personal use cases, and performance claims.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.