← BlogAugust 24, 20266 min read

Your next employee might be a bot: what Hermes Agent's update means for small business

Two Hermes Agent releases in ten days fixed two of the specific things that have made AI useless for a 40-person company. Here is what changed, and the one rule worth holding as agents get more autonomous.

If you have been putting off figuring out AI agents until things settle down, this week is a reminder that they are not going to settle down. The clearest example: Nous Research's Hermes Agent, a widely-used open-source AI agent platform, shipped two updates in the last ten days that lower the bar for what a small team can automate without hiring anyone new.

The releases themselves are not the interesting part. What is interesting is that two of the specific things that have made AI useless for a 40-person company just got fixed.

The re-explaining problem

Here is the pattern in almost every small business that has tried AI seriously. Somebody in the office gets good at it, builds up a rhythm, gets real work out of it for a few weeks. Then they go on vacation, or they move to a different project, and everything they knew evaporates, because it lived in a chat window and in their head. The next person starts from nothing.

That is what Bot Mode, launched August 17 and bundled by default from v0.20.3, is aimed at. Instead of a fresh chat every time, you build named, persistent Bots, each with its own role, model, memory, and skills. A Bot you build for quoting keeps its context, its tone, and what it has learned about your pricing every time you come back to it. Bots can hand work to each other, so one handling inbound leads can pass a qualified one to another that books the call, and small groups of up to six can work a problem across several rounds.

For a company with 25 to 150 people, that is the difference between AI as a personal trick one employee knows and AI as something the business owns. The knowledge sits in a thing you keep, not in whoever happened to learn it. If that person leaves, the Bot does not leave with them.

Scheduled work that remembers last week

The follow-up release, v0.20.5 on August 19, was reliability work rather than anything flashy, but it matters more day to day.

Scheduled jobs now keep memory between runs. Before this, an agent set to run every Monday woke up with amnesia, so a weekly competitor price check would hand you the same list of prices with no sense of what had changed, and a weekly pass over open support tickets would flag the same twelve tickets it flagged last week. Memory is what turns that from a report you skim into a report that says what is different. Nobody reads the first kind twice.

Web search also works out of the box now, rotating across five free providers, with no API key to set up first. That sounds like a footnote and it is not. The reason most owners never get past the tire-kicking stage is not the monthly fee, it is the ninety minutes of account provisioning before you learn whether the thing is useful at all. Removing that means someone can test a real recurring task on a Tuesday afternoon instead of scheduling it as a project.

What this is actually good for in a business your size

The honest version is unglamorous. Agents are good at the repeatable coordination work that eats an owner's evenings and does not require a judgment call.

Reading an inbound request and pulling out the details somebody would otherwise retype into a system. Chasing the quote that went quiet nine days ago, which nobody has time to track. Watching a customer portal or a vendor site for a change that would otherwise surprise you. Answering the internal question that gets asked forty times a year, where the answer is buried in a folder nobody can find. Producing the Monday summary that someone currently assembles by hand on Sunday night.

None of those are impressive. All of them are hours, and in a 40-person company those hours belong to someone whose time is worth more than the task. That is the whole argument, and it has not changed because Bot Mode shipped. What changed is that the setup cost dropped enough that testing it is a small decision.

Where it will bite you

Same caution as always, and it gets sharper as these things get more autonomous. On August 18, AWS made Bedrock AgentCore Payments generally available, which lets an agent discover and pay for the services it needs to finish a job, with spending ceilings set at the infrastructure layer. Read that twice as an owner. The category is moving toward agents that plan, remember, act, and spend, with less human hand-holding at every step.

The failure mode is not the one people expect. Agents rarely break and stop, which would at least be obvious. They keep running confidently and wrong, for weeks, in your name, and you hear about it from a customer. That is survivable when the job was drafting a summary. It is not survivable when the job touched a customer quote, a patient record, a certificate of insurance, or a payment.

The line worth holding is simple. An agent can draft, gather, watch, and route. A person approves anything that leaves the building or moves money, until you have watched the thing be right enough times to change your mind. That rule costs you almost nothing and it is the difference between a tool and a liability.

What to do next

You do not need to overhaul anything this week. Three low-risk steps.

Pick one recurring task you already do by hand and build a single persistent Bot for it. One. The instinct is to automate five things at once and end up with five half-working things you do not trust. Choose the task that is genuinely repetitive, where being wrong is cheap and obvious.

Test a scheduled job on something internal first, like a Monday roundup for your industry, before an agent touches anything a customer sees. You are not testing whether it can do the task. You are testing whether it is right on the days you are not watching.

Revisit in four to six weeks. This is moving fast enough that waiting a year is not a neutral choice, it is a decision to be a step behind whoever in your market is already testing.

If none of it feels urgent, that is fair. But twenty minutes with something like Hermes Agent is worth it just to see what an AI teammate that remembers things feels like. It is a different experience than a chatbot, and for most small businesses that difference is the entire point.

Sources

Want this looked at for your company?

An AI Blueprint is a short scoping session and a written plan: which of your processes are worth automating, in what order, and what each one would take. Quoted in full before anything starts.

Get your AI Blueprint