What Is Hermes Agent? A Guide for Australian Businesses

Hermes Agent is a free, open-source AI agent from Nous Research that saves what it learns as reusable “skills”, and searches for it in Australia have climbed sharply this month. Agents are in the news again this week too: on 10 November 2026, n8n Labs released Nod, a free, open-source personal agent that asks before it acts for you. Here is what Hermes Agent is, how it compares with Nod and hosted agent services, and what Australian businesses should check before letting either one near real systems.

What is Hermes Agent?

Nous Research describes Hermes Agent as “the self-improving AI agent”. It is not a chatbot that only answers questions. It is an agent harness: software that connects an AI model to tools so it can run commands, browse, work with files and carry out multi-step tasks. According to its documentation and GitHub page:

• It is open source under the MIT licence. The agent is free to run on your own computer or server. You pay separately for the AI model you connect, and Nous offers its own paid Nous Portal service as one option.
• It works with many models. You can connect Nous Portal, OpenRouter, OpenAI or your own endpoint, among others, and switch without code changes.
• It lives where you chat. You can use it in a terminal, in the Hermes Desktop app, or through messaging apps including Telegram, Discord, Slack, WhatsApp and Signal.
• It can run jobs unattended. It has a built-in scheduler for recurring tasks, can hand work to subagents, and can connect to MCP servers for extra tools.

If you are new to the terms, our guide to AI agents vs chatbots explains the difference.

How Hermes Agent learns: skills and memory

The feature behind the attention is what Nous calls a closed learning loop. The skills documentation says that when the agent works out a multi-step workflow, hits a dead end and finds the working path, or is corrected by you, it can save the approach as a skill for next time. A skill is a plain-text instruction file (SKILL.md), compatible with the open agentskills.io format, and the agent loads it only when a task needs it. Hermes also keeps a persistent memory of facts about you and can search its past conversations.

Two details matter for a business. First, the documentation says “the agent can modify or delete any skill”, so the agent is writing its own operating instructions. Second, there is a setting to require your approval before skill changes land, but it is off by default. Turn it on: a skill learned from a mistaken correction, or from a malicious web page, could otherwise be reused without anyone noticing.

Hermes Agent vs Nod vs hosted agents

Nod, announced on the n8n blog, comes at a similar job from a different direction. It is a desktop app for Mac and Windows that works in your connected apps, such as Gmail, Calendar, Drive and Slack, and on websites, inside its own sandbox and browser. It is a standalone research preview from n8n Labs, you don’t need an n8n account, and it is licensed under Apache 2.0. Like Hermes, it writes down how it got a task done and reuses those notes. Hosted agents, offered by AI vendors as a cloud service, are the third option.

Self-hosted agents give you control and keep files on your own machine, but you own the setup, updates and security. Hosted services shift that work to the vendor, at the cost of less control over where your data goes. Neither removes the need for clear limits on what the agent may do. Our page on automation vs agents covers when a simple workflow is the better choice.

Read the fine print on agent claims

Both products come with big claims, and they are the makers’ claims, not results we have tested.

• Nod’s “98% tasks finished”. This figure is being shared as if it describes your inbox. The Nod GitHub page says it comes from 226 errands run three times each, using one model (GPT-6.1 Sol) in October 2026, “in simulated accounts and websites. Test results, not a guarantee.” The launch blog doesn’t use the figure, and it says Nod’s permission check “is itself a model and can make mistakes”. The same GitHub page reports that in 240 test decisions of that check, GPT-6.1 Sol and Claude Opus 5.5 got every one right but Claude Haiku 5.5 let 8 unsafe actions through, so the model you pick matters.
• Hermes’s “self-improving”. This means the agent saves and edits its own skills and memory. It does not mean the agent gets more reliable on its own. A badly learned skill is repeated just as faithfully as a good one, which is why human review of skill changes is worth switching on.

Is Hermes Agent safe to run? Start in a sandbox

Hermes Agent is not plug-and-play. Its configuration guide says the default “local” backend runs commands “directly on your machine with no isolation”, and command-line tools keep using the credentials already set up in your normal account. Hermes checks each command against a list of dangerous patterns. By default, flagged commands are assessed by a second AI model: low-risk ones run, clearly dangerous ones are refused and uncertain ones are sent to you for approval. Commands that aren’t flagged run without asking. You can switch to always being asked about flagged commands, or turn checks off entirely with what Hermes calls YOLO mode.

The safer path, set out in the security documentation, is to run it in a Docker container, on a separate server over SSH, or in a cloud sandbox. Then a bad command damages a disposable environment, not your laptop or your business data. In those setups Hermes skips its dangerous-command check because the container is treated as the boundary, so keep real passwords, keys and customer data out of it. Prompt injection, where instructions hidden in a web page or document steer an agent, is the other risk to plan for. Our post on hidden chatbot instructions shows how much these systems depend on text you can’t see.

What to check before using Hermes Agent in your business

1. Start in a sandbox. Use a container, a spare machine or a cloud sandbox with test data, never your main work computer or production systems.
2. Give it the least access it needs. Create separate accounts and API keys with tight limits and spending caps. Our post on LLMjacking shows what leaked AI keys can cost.
3. Keep approvals on. Leave command approval on, never use YOLO mode on a machine that matters, and turn on approval for new or changed skills.
4. Know where your data goes. Whatever you give the agent is sent to the model provider you connect, which may process it overseas. Check that against your Privacy Act obligations and client agreements before using personal or confidential information.
5. Write down the rules. Decide which tasks the agent may do alone, which need sign-off, and who checks the logs. The lessons from AI agent safety in Australia apply to small agents as much as big ones.
6. Learn the basics before you build. If you are among the people searching how to build an AI agent, our free AI agents course covers how agents work and where they go wrong.

Hermes Agent and Nod are good ways to learn what agents can do. For anything that touches customers, money or personal information, an agent built with clear limits, approvals and audit trails is a safer path. That is the work of experienced AI agent developers.

This is general information based on Nous Research’s Hermes Agent documentation and GitHub page, and n8n’s Nod launch blog and GitHub page, as at 12 November 2026. Features and defaults change often, so check the current documentation before you install. Product claims are the makers’ own and have not been independently tested by Aideveloper. It is not security or legal advice for your organisation.

Featured image: Photo by Nemuel Sereti on Pexels.