What Is Hermes Agent? How a Self-Improving AI Agent Works and How It Compares to OpenClaw
Introduction
In early 2026, Nous Research released Hermes Agent — a self-improving AI agent whose tagline is "The agent that grows with you." Fully open source under the MIT license, it has accumulated over 100,000 GitHub stars as of April 2026, making it one of the fastest-growing agent frameworks of the year.
The latest version is v0.10.0 (released April 16, 2026), shipping with 70+ built-in skills and 50+ optional skills. Beyond the CLI, it supports over 15 platforms including Telegram, Discord, Slack, WhatsApp, Signal, Matrix, and more.
Key takeaways
- Hermes Agent auto-generates reusable skill files after every task via a closed learning loop, reportedly speeding up repeat tasks (no official quantitative figure has been published)
- As of April 2026, no publicly known agent-specific CVEs have been confirmed for Hermes, in contrast to OpenClaw which had multiple CVEs reported in March 2026
- Fully open source under MIT, installable via the official curl install script
What Is a Closed Learning Loop?
The closed learning loop is the core of Hermes Agent. After completing a task, it analyzes the steps taken and automatically generates a reusable skill file. The next time a similar task arises, that skill is applied immediately.
This loop consists of three elements:
- Automatic skill generation: After task completion, steps are analyzed and a reusable skill file is created. Each use refines the skill further
- Self-evaluation cycle: Every 15 tasks, the agent evaluates its own performance and identifies areas for improvement
- Cross-session search: FTS5-powered cross-session search with LLM summarization enables efficient retrieval of past knowledge
Self-created skills are reported to speed up repeat tasks, although no official quantitative figure has been published at this time, so real-world gains will depend on your specific workflow. This is not just a memory feature — it is a built-in "learn and improve" cycle.
On top of this, agent-curated memory with periodic nudges and Honcho dialectical user modeling enable continuous learning of user intent and preferences.
The Gateway and Agent Two-Tier Architecture
Hermes Agent separates its runtime into a Gateway layer and an Agent layer. This split is what makes "serverless operation + scheduled tasks + persistent messaging integrations" possible in one system.
Gateway layer (lightweight, always-on)
- Connects to 15+ messaging platforms (Telegram, Discord, Slack, etc.)
- Built-in scheduler that ticks every 60 seconds to fire cron jobs
- Session routing and webhook reception
- Lightweight enough to run on a minimal VPS ($1-5/month)
Agent layer (heavy compute, serverless)
- LLM calls, browser automation, shell command execution
- Skill and memory read/write against a persistent volume
- Spins up only when a task fires, hibernates afterward
- With Modal or Daytona, cost drops to zero during idle periods
For example, "every morning at 8am, check Gmail and Slack for important messages and summarize" is parsed by the LLM into a cron expression (0 8 * * *) and stored in the Gateway's internal database. At 8am, the Gateway triggers a new agent session on Modal, the task runs, and the session hibernates. You only pay for the few minutes the agent actually ran.
Modal and similar AI-focused serverless platforms provide persistent volumes, so skill files and memory databases survive across sessions. Stateless services like Cloud Run don't have this — you'd need to wire up external storage yourself, which makes them a poor fit for Hermes.
How Does It Compare to OpenClaw and Claude Code?
Hermes Agent occupies a distinct niche among competing agents. In short, Hermes focuses on deep learning loops, while OpenClaw emphasizes broad integrations.
Hermes vs. OpenClaw
| Aspect | Hermes Agent | OpenClaw |
|---|---|---|
| GitHub Stars | over 100,000 | over 350,000 |
| Core Strength | Self-improving learning loop | Broad integrations |
| Security | No publicly known agent-specific CVEs as of April 2026 | Multiple CVEs reported in March 2026 (GitHub Security Advisory) |
| UX | Still maturing | Mature |
| License | MIT | Apache 2.0 |
OpenClaw had multiple security vulnerabilities (CVEs) reported in March 2026 (see GitHub Security Advisory). Hermes Agent, by contrast, has no publicly known agent-specific CVEs as of April 2026.
Hermes vs. Claude Code
Claude Code is known for its strength on coding tasks. Hermes Agent has not published independent SWE-Bench scores — its focus is general-purpose autonomy rather than coding-specific performance.
Using Both in Practice
An interesting real-world pattern is using OpenClaw as the orchestrator and Hermes as the execution specialist, connected via the ACP protocol. Rather than strict competitors, the two are complementary.
Installation and Setup
Getting started with Hermes Agent takes minutes. The setup is interactive and straightforward.
Install
The officially recommended method is to run the curl install script. It supports Linux, macOS, WSL2, and Termux (Android).
# Official install script (Linux/macOS/WSL2/Termux)
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bashThe installer handles dependencies (Python, Node.js, ripgrep, ffmpeg), clones the repository, sets up a virtual environment, and configures the global hermes command automatically.
Setup and Launch
# Initial setup (includes OpenClaw migration wizard)
hermes setup
# Launch
hermes
# Launch in TUI mode
hermes --tuiSupported Models
Hermes Agent works with a wide range of LLM providers:
- Nous Portal / OpenRouter (200+ models)
- NVIDIA NIM / Kimi/Moonshot
- OpenAI / Hugging Face
- Custom endpoints
A minimum context window of 64K tokens is required.
Deployment and Pricing
Hermes Agent supports 6 terminal backends: local, Docker, SSH, Daytona, Singularity, and Modal. Daytona and Modal offer serverless persistence — the agent's environment hibernates when idle and wakes on demand.
With serverless deployment, costs between sessions drop to nearly zero. Unlike an always-on VPS ($5-10/month), you pay only for the time the agent is actively running.
The framework itself is free under the MIT license. LLM API costs run approximately $0.30 per complex task on budget models. Pick local, VPS, or serverless depending on your workload.
Why Security Matters
Security is a major differentiator for Hermes Agent. As of April 2026, no publicly known agent-specific CVEs have been confirmed.
OpenClaw had multiple CVEs reported in March 2026 (see GitHub Security Advisory). Given that agents operate with broad permissions in a user's environment, this is a risk that cannot be ignored.
As a newer entrant, Hermes Agent incorporated security-first design from the outset. For enterprise environments evaluating agent adoption, CVE history is a critical decision factor.
FAQ
Q. Is Hermes Agent free to use?
A. The framework itself is fully free under the MIT license. However, LLM API usage incurs separate costs. With budget models, complex tasks cost approximately $0.30.
Q. Can I migrate from OpenClaw?
A. Yes. The hermes setup command includes a built-in OpenClaw migration wizard that helps transfer your existing configuration and workflows.
Q. Is it a coding-only tool?
A. No. It supports over 15 platforms — CLI, Telegram, Discord, Slack, WhatsApp, Signal, and more. It handles general tasks beyond just coding.
Q. Which model should I use?
A. Hermes supports over 200 models via OpenRouter, plus Nous Portal, NVIDIA NIM, Kimi/Moonshot, OpenAI, and Hugging Face. A minimum context window of 64K tokens is required.
Q. What about SWE-Bench scores?
A. Hermes Agent has not published independent SWE-Bench scores, as it focuses on general-purpose autonomy rather than coding-specific benchmarks. If coding performance is a priority, consider using Claude Code or OpenClaw alongside Hermes, or evaluate them in combination.
Summary
Hermes Agent brings a fundamentally different approach to the AI agent landscape. Instead of competing on breadth of integrations or coding benchmarks, it bets on self-improvement — a closed learning loop that makes the agent measurably better with each task.
No agent-specific CVEs reported, full MIT licensing, and support for 15+ platforms make it a compelling option for enterprise adoption. At ZenChAIne, we continue to track the evolution of AI agents and share practical insights on how to put them to work.