Your AI Agents Can Now Talk to Each Other. Here's What to Automate First.

For two years, your AI assistant has been a genius that can only do one thing at a time. That just changed.

It is 6 p.m. In Manchester. Priya Mehta has just climbed out from under a customer's sink, and her phone will not stop buzzing. Three quote requests came in while she was elbow deep in pipework. Two invoices from last week are still unpaid. A customer from Tuesday is asking, again, whether her callout charge is correct.

Priya runs a six-person plumbing and heating firm. Right now, she is also the entire front office. The quoter. The chaser. The bookkeeper. The one person who remembers who asked what, and when.

Here is what that actually looks like, most weeks:

• A lead texts. Priya replies when she gets a minute, sometimes six hours later.
• A quote goes out. Nobody follows up on it unless Priya remembers to.
• An invoice sits unpaid until Priya happens to notice, usually a week too late.
• A customer question lands in a text, a missed call, and an email, and Priya is the only one checking all three.

She is not running a business. She is running four jobs badly, alone, after hours.

AI agents just learned to hand work to each other

On 13 August, Teknium, the engineer who leads Hermes Agent at Nous Research, posted three words that mattered more than they looked: Bot Mode is here. Instead of one long chat with one AI profile, Hermes now lets you set up a roster of named bots, each with a job, a short description of what it does, and its own avatar. They message each other directly. One bot can hand a task straight to another without you relaying it in the middle.

The post went out to Teknium's own following and, within eight hours, had been seen more than 265,000 times, drawn nearly 2,000 likes, and been saved over 1,000 times by people planning to actually build with it. That is not hype cycle noise. That is a lot of builders recognising, at the same moment, that something changed.

It is not an isolated move, either. Days earlier, Cursor and xAI shipped Grok Bot with almost the identical idea: name an agent, give it a title and a description, and it will recognise when a task belongs to a different agent and quietly pass it along. Nate Herk, who teaches AI automation to a large audience of builders, spent twenty minutes walking through exactly how that handoff works in practice. His two named agents, Klaus the executive assistant and Dev the engineer, message each other mid-task without a human relaying anything. When he asked Klaus to build a waitlist page, Klaus looped in Dev on its own.

Two competing labs shipping the same idea, in the same week, is the signal. This is not a gimmick either company invented alone. It is the next obvious step once agents got good enough to be trusted with a job description instead of a single prompt.

One assistant that does everything, or three that each do one thing well?

Most small businesses that have tried AI so far have tried it Priya's way: one chatbot, one inbox, one long conversation that is supposed to cover quoting, chasing, and reporting all at once. It works for a week. Then the conversation gets long, the AI starts mixing up which customer is which, and everyone quietly goes back to doing it by hand.

A named roster fixes the actual problem, which was never intelligence. It was scope. Ask one assistant to be your receptionist, your collections officer, and your analyst, and it will do all three averagely. Give three agents one job each, a name, and a short description of what they are for, and each one gets sharper, because it is not trying to hold everything in its head at once.

Isn't this just three chatbots pretending to be a team?

Fair question, and worth asking before you build anything. The honest answer is: mostly no, with one real caveat.

The difference from a normal chatbot is the handoff. A chatbot answers what is in front of it and stops. A named agent with a job description actively looks at an incoming task, checks whether it matches its own job or a teammate's, and routes it, the same way Nate Herk's Klaus looked at a build request and looped in Dev without being told to. That routing is the actual new capability, not the fact that the agent can chat.

The caveat: a bot roster is only as good as the job descriptions you give it, and it can still misroute an edge case. Start narrow, watch it for a week, and keep a human approval step on anything involving money leaving your account. Every platform doing this, Hermes Agent included, ships an approval setting for exactly that reason.

What is Hermes Agent's Bot Mode, exactly?

Bot Mode replaces Hermes Agent's old "sessions" view, where every agent profile lived in its own separate chat that you had to open one at a time. Under Bot Mode, each profile becomes a named bot with a title, a short job description, and a profile picture, all visible together in one sidebar, like a team roster instead of a folder of chats. Bots can message each other directly through an inbox, be assigned recurring jobs such as cron-scheduled reports, and be reviewed individually so you can see exactly what each one did and when. Nous Research opened it as a one-day public beta before a wider rollout, which is part of why the reaction on X was so immediate.

The underlying capability, agents that can hand off to each other instead of routing everything through one conversation, is sometimes called agent-to-agent communication, or A2A. It is the same idea Grok Bot shipped as "agents messaging each other," and the same idea a growing list of developer frameworks (AutoGen, CrewAI, LangGraph) have offered to programmers for over a year. What changed this week is that it arrived with no code required, inside a product a non-technical business owner can actually open and use.

How to build your first bot roster this week

You do not need Priya's exact business for this to apply. Swap in your own three bottlenecks: quoting, collections, inventory, scheduling, reviews, hiring, onboarding, whatever eats your evenings. The steps stay the same.

1. Write down your three worst bottlenecks, not twelve. Priya picked customer intake, lead follow-up, and weekly reporting. Pick the three tasks that already live in text, email, or a spreadsheet, because those are the easiest to hand to an agent first.
2. Give each bot a name, a title, and one sentence describing its job. Not a paragraph. "This bot answers the first message from a new lead and books a callback if I am not free" is enough for the router to work with.
3. Connect the shared tools once. Email, calendar, WhatsApp or SMS, whatever you already run the business on. In both Hermes Agent and Grok Bot, connecting a tool to one bot makes it available to every other bot on the roster, so you are not re-authorising the same login three times.
4. Test the handoff on a fake job before a real one. Send your Intake bot a message pretending to be a slow-paying customer and see whether it correctly flags the Follow-Up bot instead of trying to close the invoice itself.
5. Turn on approval mode for anything involving money or a customer-facing message going out unsupervised. Let the bots draft and route freely. Keep yourself as the final click on anything that leaves the building.
6. Review what happened at the end of week one, not year one. A named roster produces a clean log of who did what. Read it. Tighten the job descriptions that caused confusion, and leave the ones that worked alone.

Priya's version, once she tried it: an Intake bot that answers a new lead within minutes and books a callback slot. A Follow-Up bot that checks, three days later, whether a sent quote has been accepted, and nudges the customer if not. A Report bot that reads both of their activity every Friday and sends Priya a two-line summary of what needs her attention Monday morning. None of it required a developer. All of it ran on tools she was already using for email and messaging.

What you have after this

A week in, you have three named agents doing one job each instead of one tired assistant trying to do four. A month in, you have a log of every lead that got a same-hour reply instead of a same-week one, and a follow-up that never once forgot to chase a quote. Six months in, you are not managing software anymore. You are managing a small team that happens to be made of agents, each one doing exactly the job you gave it, quietly handing off what is not theirs.

Same business. Same customers. A completely different front office.

If you want help mapping your own three bottlenecks into a working bot roster without weeks of trial and error, that is exactly what Notma's 14-Day AI Readiness Sprint is built for. You can also read our earlier guide, What Is Hermes Agent, and Should Your Business Use It?, for the basics before you add a second or third bot to the roster.