In the first piece in this series I said that a model on its own does not do very much, and that to get real work out of it you need a second thing, a piece of software that runs the model on a schedule, gives it tools and memory, and lets you talk to it. That layer is called an agent runner, and the open one worth knowing is Hermes, from a lab called Nous Research. This is the plain-English guide to what it actually does, because once you strip away the science-fiction language, an agent runner is less of a robot and more of a tireless junior who quietly works through the boring jobs while you get on with the things that need a human.
From a chatbot to a colleague: what an agent runner is
The difference between a chatbot and an agent is the difference between a tool you pick up and a colleague you delegate to. A chatbot waits for you to type a question, gives you an answer, and then forgets all about it. An agent runner keeps a model running in the background on a machine you control, often something modest like a spare Mac Mini left switched on in the corner, and it lets that model do things rather than just answer, which means reading and writing files, searching the web, calling other software, remembering what it has done before, and waking itself up on a timetable to get on with jobs without being asked. Hermes is one of these runners, it is open-source and free to install, and the thing that makes it flexible is that you can point it at whatever model you like, whether that is a paid one such as Claude or a cheap and open one such as the GLM-5.2 from the first piece, and you talk to it through an ordinary chat app rather than a terminal.
The words you need: skills, cron and the night shift
A handful of terms come up constantly here, and they are all simpler than they sound. A skill is just a written instruction sheet that teaches the agent how to do one repeatable job well, the digital equivalent of a standard operating procedure, so a morning-briefing skill might set out where to look, what to summarise and how to lay it out, and the agent reaches for it only when that particular job comes up. A cron job is the part most people ask about, and it is simply a task set to run automatically on a timetable, so "every weekday at eight" or "every Wednesday at ten" is a cron job, and although the odd little word comes from an old Unix scheduling tool, the friendliest way to picture it is as a digital alarm clock that wakes the agent to do a set piece of work.
The background jobs, or the "night shift" as the people who run these setups tend to call them, are exactly what they sound like, the work you leave running unattended while you sleep, so that a finished summary or a tidied set of files is waiting for you in the morning. The way you actually talk to the agent is through a chat app, usually Discord, Telegram or Slack, and most experienced users settle on Discord because it lets you keep separate channels for different parts of the business and a proper thread for each job, which is far easier to live with than one long scrolling conversation.
What people actually hand to it
The work people delegate to these agents is far more ordinary than the hype suggests, and that is exactly why it is useful. The most common job is the recurring summary, where the agent wakes each morning, gathers the news or the numbers you care about, and drops a short briefing into your chat before you have made the coffee. Close behind is capture and filing, where it quietly takes in everything you produce, the documents, the transcripts, the published posts, and files them in a tidy structure so that nothing is lost and everything can be found again later. Beyond those, the same patterns turn up across very different businesses, from weekly checks on how a website is performing, to a first pass over a developer's code before a human signs it off, to research and background briefs prepared ahead of a meeting, to the small operational chores that are too fiddly for proper software but too repetitive to keep doing by hand, such as watching payments land and summarising the week's revenue. None of that is a robot running your company, and that really is the point, because the value is in quietly handing over the dozens of small, dull, scheduled jobs that otherwise eat a working week.
Don't marry a platform
If there is one strategic point to take from all of this, it is that you should not fall in love with any single one of these tools, because the platforms themselves are changing every few weeks, with one fashionable runner giving way to the next before you have finished learning it. The people who do this well treat the runner as disposable and keep the valuable part portable, and the valuable part is never the software, it is the skills, the instructions and the patterns you have worked out, which are simply written files that you own and can carry from one runner to the next. There is even an emerging shared standard for these skill files, and Hermes can import another runner's setup wholesale, which tells you that the people building these tools already assume you will want to move. This is the same lesson the first piece reached from the other direction, that the sensible posture is to stay loose and governed rather than locked in, whether the thing you might be locked into is the model or the software that runs it.
Where it can bite you
It is not all upside, and there are a few things worth going into with your eyes open. The first is a small but genuinely useful thing to understand about how these scheduled jobs work. When a job runs on a timer it does not carry over the casual back and forth you had with the agent in chat, so you cannot set one up by telling it to do "that thing we talked about last week" and expect it to know what you mean. Everything the job needs has to live in its skill, the written instruction sheet it loads each time it runs, which is exactly why a well-built scheduled task is reliable: the skill is the memory, not the conversation. That is also the honest weakness of unattended work right now, because a job is only ever as good as the instructions you have written into its skill, and these tools are still new enough that jobs do quietly fail and need an eye kept on them.
The bigger thing to be careful about is security, and it comes from the very feature that makes an agent useful in the first place. An agent that can see your files, read material from the outside world such as emails and web pages, and then send information back out has, all together, everything an attacker would need, because a booby-trapped email or web page can quietly tell the agent to hand over your data, and the developer Simon Willison has neatly named this combination the lethal trifecta. The defence is not complicated, it is the same common sense you would apply to a new member of staff, which is to start the agent off read-only on test accounts, keep it walled off from anything precious, and make it check with a human before it ever sends, deletes or pays for anything.
Two last things are worth deciding on purpose rather than by default. One is which model to put behind a job, because the routine, high-volume, low-stakes work runs perfectly well on a cheap or local model, while the jobs where a mistake is expensive should go to a frontier model such as Claude and stay under a human eye. The other is worth checking before you rely on a paid model inside one of these runners, because the rules on using a monthly subscription this way have been changing lately, so for anything serious it is usually cleaner to use a metered, pay-as-you-go key than to assume your plan will cover it.
What this actually changes
What has genuinely changed is that this is now within reach of an ordinary small business rather than only a software team, because a runner like Hermes can be set up in an afternoon and pointed at whichever model suits the job. It still rewards exactly the discipline you would bring to any new system, which is to start small, keep a person in the loop on anything that matters, and own your patterns rather than the platform. That last point, owning the discipline rather than the tool, is where this quietly turns into a governance question, because every one of those scheduled jobs leaves a record of what was done, and that trail is either a liability or one of the most useful assets you have, depending entirely on how you handle it. That is where the final piece in this series goes next.
Sources
- Hermes Agent, the open-source runner: Nous Research.
- The "lethal trifecta" for Ai agents: Simon Willison.
- The open, portable Agent Skills standard: agentskills.io.
- Real-world agent-runner usage patterns: Builder Methods, and the deep-research briefings gathered for this series.