Hermes Agent is good at recurring work: jobs that run on a schedule, handed out in the chat app your team already uses, by an agent that gets better at them over time. It is not something you install and forget. We have run our own team of 13 agents on it since early July 2026, and our verdict is that it is worth it when one named person owns its upkeep.
Two disclosures before the scores. We sell a done-for-you service built on Hermes Agent, so this is a review from people who liked it enough to bet a business on it. And since August we have run our own maintained version, built from Nous Research's code and updated on our own schedule. Every statement below about what the software does comes from the official documentation or the project's public page as read on October 26, 2026, so it describes the version you would download.
Scores are out of 5. They are our judgment from our own records rather than a benchmark.
| What your business needs from it | Our score | What we saw |
|---|---|---|
| Doing jobs on a schedule | 5 | 118 scheduled jobs switched on across 12 agents today |
| Working from your team's chat app | 5 | Each of our agents is its own named teammate in Slack |
| Learning how you do things | 4 | It writes its own procedures, but a written lesson is not a fixed habit |
| Remembering the right things | 3 | Its notes are short on purpose, and some turn out wrong |
| Holding back when nobody is watching | 4 | Risky steps in scheduled jobs are refused by default |
| Updates | 2 | Eight releases in 16 days this month, and someone has to apply each one |
| Getting started without technical staff | 2 | Installing takes minutes; defining the work is the real project |
Is Hermes Agent worth it for a small business?
Yes, for recurring work that needs some judgment, if someone on your side will own it. The software is free under the MIT license. What you pay for is the person who keeps it on track. If nobody can take that on, have it run for you or wait.
That person does not need to be an engineer for the daily work. They need to read what the agent produced, notice when it is wrong and fix the instruction or the note behind the mistake. When an update breaks something, though, they will need help from someone who is comfortable with software, or a service that handles it.
What it costs in money is small next to that. The model, the machine and the hours are broken into five lines, with a worksheet, in what Hermes Agent costs a small business to run. If you are still getting your bearings on what the software is, start with our plain explainer on Hermes Agent.
What does Hermes Agent do well?
Scheduled work is where it is strongest. On October 26, 2026, 12 of our 13 agents had scheduled jobs switched on, 118 in all, ranging from checks every few minutes to monthly reports. The first one we set up dates from July 5. It also works well as a teammate in chat, and it gets better at tasks it repeats.
Hermes also lets a scheduled job skip the AI entirely. Its scheduled jobs guide describes jobs where a plain program runs on schedule and its result goes straight to your chat app. Forty-one of our 118 jobs work that way. A check that only needs to compare two numbers should not pay for an AI model to read it, and it behaves the same way every time. If a good share of your list looks like that, see whether a plain automation tool fits that part better.
The chat side works the way an owner would hope. Each of our agents has its own name in Slack and its own logins, memory and skills, and our team messages them like coworkers. Hermes connects to more than 20 chat apps, Microsoft Teams, Telegram and WhatsApp among them. Its desktop app now offers the same one-agent-per-job setup as Bot Mode, where, in the documentation's words, "a Bot is a Hermes profile" with its own memory, skills and logins. Keeping each agent's access separate is the single choice we would make again first, because an agent that does everything needs access to everything.
The learning is real. After a job, Hermes writes up how it did the work as a reusable procedure, which it calls a skill, and it improves that procedure on later runs. A built-in tidy-up called the curator is on by default. It marks the skills the agent wrote as stale after 14 days without use and archives them after 30, and it never deletes them outright, so the list does not fill up with near-copies.
Where does Hermes Agent fall short?
Its memory is small and can be wrong, a written lesson does not always change its behavior, and its own report of "done" is a claim you still have to check. None of them is a reason to avoid it, but each one costs time every week that no feature list mentions.
Its memory is short, and some of it will be wrong
Hermes keeps two small notebooks. Its documentation gives the limits: 2,200 characters for what the agent has learned and 1,375 for what it knows about you, which is a few short paragraphs each. That is deliberate, since the notes hold key facts and searchable past conversations hold the rest. By default the agent saves notes on its own. You can turn on approval so that every note waits for a person.
The trouble is that whatever an agent writes down stays trusted after it stops being true, and some of it is wrong from the start. On October 14 our publishing agent added a line to its own instructions saying our website had no separate field for the title Google shows. It had read an out-of-date copy of the site. The field existed, and the line was corrected the same day. Our standing rule is to fix a wrong note or instruction the moment anyone spots it, without waiting to be asked. Budget a few minutes a week per agent for reading what it believes.
A written lesson is not the same as a fixed habit
From October 7 to October 14, our publishing agent made the same small mistake on five days out of eight. It stamped an article with a release time a few minutes in the future, and the website quietly leaves out anything dated after the current time. It caught the problem each time before the article went out, so no reader saw it, but it cost a rebuild every time.
The agent recorded each slip in its own procedure, with the dates, and it kept happening. What has held since is a mechanical change. The agent now leaves the release time as a blank marker that the site refuses to accept, and fills in the real time only at the moment of release. We have not logged the mistake since October 14.
If an agent repeats a mistake, change the check that should catch it. Rewriting its notes did not work for us.
Its report of "done" still needs checking
Hermes's own security documentation says that when it blocks a change to a file, "The model may still claim the edit succeeded." It tells you to trust its verification note over the agent's closing summary. Our own record says the same: the first of the five failures we wrote up is an agent reporting success while the work was gone. We now judge a job on whether the file, the draft or the page exists.
What does it take to keep Hermes Agent running?
It takes a named owner who checks the output and plans for updates, and updates are the part people underestimate. The project shipped eight releases between October 8 and October 23, 2026, and on October 26 its public page listed 17,228 open issues and 33,971 proposed changes waiting for review. Whoever owns your agent will run into open bugs and will have updates to apply most weeks.
The official updater gives you some protection. Before each update it saves a copy of each agent's settings and scheduled jobs, and it pauses running agents while the update is applied. The documentation is clear about the limit: "Quick snapshots recover state files, not application code." The software itself is not covered by that copy.
We found out why that matters on the morning of August 13. A change was applied to the copy of the software our nine agents were running at that moment. It rewrote about 1,900 of its files underneath them, and all nine broke together. Since August 22, each new version is installed alongside the one that is running, we switch over in a single step, and the previous version stays on the machine so we can go back.
At our size, the software itself is one agent's job. Owen ships the new versions our own agents run on, and he also sets up our agent product for customers. With one agent you will not need a full role, but you will need a named person and a way back.
Can you trust it with work nobody is watching?
Mostly, because Hermes is careful by default when no one is there to approve. Its security page lists eight layers. The one that matters most for a business is that a scheduled job that hits a risky step is refused instead of approved, and a request for approval that nobody answers in time also counts as a no.
We add our own rules on top. An email from an owner's mailbox to anyone outside the company waits until a person approves that exact draft, and nothing that moves money happens without an owner's approval. Those limits come from what each agent can reach, which is the subject of what access an AI agent should have to your business systems, and from where a person signs off.
Who should not run Hermes Agent?
Skip it, or have someone run it for you, in four situations:
- The job is one fixed set of steps at high volume with no judgment in it. A workflow tool is cheaper and more predictable.
- Nobody on your team can own it for a few hours a month after the first month.
- Several people need different limits inside the same agent. Hermes and OpenClaw both answer that with separate agents, which our Hermes Agent vs OpenClaw comparison covers.
- The work cannot be undone if the agent gets it wrong, and no person will approve it first. That is a job to keep, as with the five we took back from our own agents.
If you are still leaning toward running it yourself, answer these five questions before you install anything:
- Who applies updates, and how will you go back to the last version if one breaks something?
- What is the one job it does first, and how will you check that the output exists?
- What must it never do without a person approving it?
- Who reads its notes and corrects them, and how often?
- Where do approval requests go when nobody is at a desk?
If you cannot name a person for the first two, wait, or have it done for you.
Frequently asked questions
Is Hermes Agent good?
Yes, for recurring work handed over through a chat app by a team that will check the output. It is weakest on upkeep. Updates are frequent and its notes need tending, so it suits a business with one person who will own it.
Is Hermes Agent better than OpenClaw?
For most owners the difference matters less than who keeps it running. Hermes leans on learning and OpenClaw on strict rules, and we compare them point by point in what actually decides Hermes Agent against OpenClaw. If you want to compare more options, our list of OpenClaw alternatives includes Hermes and seven others with their prices.
How long have you used Hermes Agent?
Since early July 2026. Our first scheduled job dates from July 5, and on October 26 we had 13 agents and 118 scheduled jobs switched on. The public project dates from July 2025.
Is Hermes Agent free?
The software is, under the MIT license, including for business use. You pay for the AI model it uses, the machine it runs on and the time it takes to keep it working.
Do you need to be technical to run Hermes Agent?
Not to install it. There is a desktop app for Windows and Mac, and the install takes minutes. Day-to-day use takes someone who will read the output and correct the agent. Fixing it after a bad update takes someone comfortable with software.
If you want it without the upkeep
We build the agent around one job in your business, connect it to your tools and keep it updated, on the same software our own company runs on. The build starts at $1,000 one time, and hosting and maintenance start at $200 a month, which covers monitoring and routine updates. Bring the job you would hand over first and start a conversation with us.
