Hermes Bot Mode organizes specialized agents into three interaction patterns: you talk to one bot, one bot messages another, or several bots work inside a capped group chat. The group chat cap is what stops a loop from draining your token budget, and the tool list is what stops a bot from doing something you would regret.
What Hermes Bot Mode Actually Changes
Hermes Bot Mode is a coordination layer for Hermes, the open-source agent released by Nous Research, and its core change is that bots can address each other directly rather than only through you. A bot in Hermes is a profile, and every profile carries its own persona file, model, tool list, MCP connections, and memory, so the two terms describe the same object. The underlying capability existed before bot mode; what is new is the ability to hand one bot a job, watch it split that job across teammates, and receive one consolidated result back.
Before bot mode, the friction was never the bots themselves, it was the coordination between them. You were the router: copying output from one chat into another, remembering which agent held which context, and manually deciding who did what next. Bot mode removes you from that loop. One orchestrator receives the job, breaks it into pieces, tags the bots who own those pieces, and reports back to you alone. If you want to see the mechanism in action before reading further, the source video walks through a full build.
The Three Ways Hermes Bots Can Work Together
Hermes bots coordinate in three distinct patterns, and each one has a different cost and oversight profile. You talk to a single bot, one bot messages another bot directly, or several bots work inside one group chat thread with a cap on how many times they can respond to each other. That cap is the part worth paying attention to, because an uncapped agent thread is an uncapped token bill.
| Pattern | Who talks to whom | Oversight | Best for |
|---|---|---|---|
| One-on-one | You and a single bot | You see every message | One-off jobs, quick edits |
| Direct | Bot to bot, no group thread | You review after the handoff | Two-step pipelines |
| Group chat | Several bots in one capped thread | You watch or step away | Recurring multi-role workflows |
The group chat is the pattern that changes how you work. In the walkthrough that produced this article, the orchestrator created the group chat itself, then delegated with explicit tags: at-researcher for the bot gathering evidence, at-creator for the bot turning that evidence into a post, and at-operator for the orchestrator who quality-checks the final package. Tagging works the same way for the whole team at once.
Creating a Bot: Name, Title, and the Description That Becomes Its Soul
Creating a bot in Hermes takes three inputs: a name, a title, and a one-or-two sentence description of the job. The title renders in the sidebar, the name is what you use to tag the bot with an at symbol, and the description is used to generate the bot's personality and operating rules. You can clone from your default profile so a new bot inherits your existing setup, then assign a model. The generated personality file, which Hermes calls the soul, is editable, and tightening it with explicit rules produces more predictable behavior than leaving the generated version untouched.
The soul file is the part most people underuse. In the recorded build, the orchestrator wrote a soul for its researcher that included a mission statement, a reporting line, and a list of responsibilities, which is more structure than a hand-written prompt usually gets. A sales bot in the same setup carried rules like never discussing prices without approval and always handing results back to the main Hermes instance rather than straight to the operator. The tighter the job description, the more reliably the bot stays inside it.
One caveat matters here: the soul is a soft control. The model tries to honor what is written in it, but nothing enforces it. Treat it as reinforcement, never as your only lock.
How Tools, Skills, and MCPs Drive Per-Message Token Cost
Hermes splits capabilities into three categories with very different cost behavior, and the distinction determines whether a bot team stays affordable. Skills are saved workflows that load only when the bot invokes them, so leaving them enabled costs almost nothing at rest. Tools are live actions such as reading a file, searching the web, or sending a message, and they load into every message the bot processes. MCPs are external integrations connecting Hermes to services such as Gmail, Notion, or GitHub, and they also load per message.
The cost asymmetry is the whole point. Every new bot starts with the same tools and MCPs your main instance has connected, all switched on, and some tools load upwards of 50,000 tokens into every single message. That figure is per message, not per session, so a chatty bot with heavy integrations pays that overhead on every turn. Disabling a tool a bot never needs makes it faster, cheaper, and easier to reason about, and it is the single step that most improves a mediocre bot.
The second effect of tool selection is capability, not cost. MCPs act as specialist powers: a sales bot wired into a CRM can edit, search, and sort leads that no other bot can see or needs to. Adding that connection once makes it part of that bot's kit for every future run.
Approval Gates: Where to Put the Line a Bot Cannot Cross
Approval gates are the difference between a bot team you supervise and one you can leave running, and Hermes gives you three places to set them. The strongest gate is simply withholding the tool: a bot without email-sending capability cannot send an email regardless of what it decides mid-task. The second is the built-in approval setting, which can be configured to run safe actions automatically and stop for a decision on anything ambiguous. The third is the soul file itself, where you write a rule in plain language such as never sending anything to a client without approval.
The three gates differ in strength, and that difference decides where each rule belongs.
| Gate | Strength | Who enforces it |
|---|---|---|
| Withheld tool | Hard | The system, no tool, no action |
| Approval setting (smart mode) | Hard | Hermes, blocks and asks |
| Soul file rule | Soft | The model, best effort |
When the approval setting stops for a decision, you get four choices: allow once, allow for the session, always allow, or deny. That makes the middle layer practical for repeated tasks, because a safe pattern you approve once can be approved for the session instead of interrupting every run.
The useful exercise is to write your own list before you build anything. In the walkthrough, the operator named three actions worth a hand on: spending money, sending something to a client, and deleting files. Everything else ran without approval. Your list will differ, but the question that produces it is the same one: what could a bot do that you would actually regret?
Two Settings That Keep a Long-Running Team Alive
Two configuration values decide whether a team survives a long job or dies halfway through it, and both take seconds to change. The first is how many bots stay warm. The default keeps three bots ready, and raising it to five means switching between bot chats is instant instead of waiting for a reboot. The second is the idle timeout, set to 10 minutes by default. If your workflows regularly run longer than that, raise it to 20 or 30 minutes depending on how much time the job typically takes, so a long task does not error out mid-run.
There is also a memory setting worth knowing about. If you run a shared memory provider such as Honcho, Memo, or Super Memory, let only the orchestrator write to it, and set every other bot to built-in memory only. The orchestrator already saves what matters to shared memory, so letting workers write as well duplicates the same facts and burns context for no gain.
A stalled team is not always a configuration problem. In the recorded session, the group chat went quiet for about five minutes and the cause turned out to be the model provider, GPT, hitting a session limit. Resetting it brought the team back. If your bots stop responding, check the provider before you start rewiring tools.
Planning Your Own Hermes Bot Team
The right team shape follows from the Stable Diffusion the task, and the tool split follows from the shape. A one-off task does not justify a team; a single competent agent handles it. Recurring weekly work such as content production, client intake, or inbox triage is where a small set of specialists pays for itself.
The recorded example used three bots: an orchestrator acting as the operator and point of contact, a researcher whose only job was finding what currently performs in a specific niche, and a creator who turned research into posts. That is enough structure to cover research, production, and review without adding coordination overhead you have to manage yourself. The operator, not you, spoke to the other bots.
Model assignment follows the same logic as tool assignment:
- Put a strong, capable model on the orchestrator, because it makes the calls and handles complex reasoning.
- Put cheap or local models on the workers, which run simple, well-specified steps.
- Only pay for the smart model where a decision actually depends on it.
That split is what keeps the whole team cheap rather than pricey. Running workers on a local model on your own machine costs nothing, and for narrow jobs you get effectively the same result.
How a Hermes Bot Team Compounds Over Time
Once a job wraps, you can tell the team to save what it learned from that run. The next time you ask for the same kind of work, the team already knows your format, your style, and the sources you trust, and it leans on you less. This is the part that compounds: you are not just running a team, you are training one, and the questions it asks you get fewer over successive runs.
Multiple workflows stay organized through groups. You can create a group, name it after the project, and move the bots that belong to it inside, so several content pages or client workflows remain navigable instead of collapsing into one long list of chats.
Common Questions About Hermes Bot Mode
- What is Hermes Bot Mode? Hermes Bot Mode is a coordination feature in Hermes, the open-source agent from Nous Research, that lets specialized bot profiles message each other and hand off work inside a shared group chat. It also supports one-on-one conversations with a single bot and direct bot-to-bot messaging.
- Is Hermes free and open source? The Hermes agent repository is published publicly, and you can check its current license and release status on the canonical project page. Because Hermes connects to external model providers, your real cost depends on which models you configure and which tools you leave enabled for each bot.
- How do you keep a Hermes bot team from getting expensive? Put the strongest model on the orchestrator that plans and coordinates, and cheap or local models on the workers that execute simple steps. Then disable every tool and MCP a bot does not need, because each enabled integration loads into every message, and some load upwards of 50,000 tokens per message.
- Can Hermes bots act without your approval? The approval setting can be configured to run safe actions automatically and stop on ambiguous or dangerous ones. Withholding a tool entirely is the strongest gate, and a rule written into the bot's personality file is the softest.
- What is the soul file in Hermes? The soul file is the bot's generated personality and rules document. You can edit it to add explicit constraints, and it reinforces behavior, but it should not be your only safeguard against an action you would regret.
- Why does a Hermes bot chat go quiet or error out? The most common causes are the model provider hitting a session limit and the idle timeout being too short for the job. Resetting the provider or raising the timeout from 10 minutes to 20 or 30 usually resolves it.
- Should a bot team share one memory? Only the orchestrator should write to a shared memory provider. Set worker bots to built-in memory only, otherwise the same information is saved twice and wastes context on every run.
- How many bots should stay warm? Three by default. Raising it to five keeps more bots ready so switching between them is instant rather than triggering a reload.
- What is the first job worth giving a bot team? Pick work you repeat every week, such as content production, client intake, or inbox triage. One-off tasks are cheaper and simpler with a single well-configured agent.
Turning a Video Walkthrough Into a Written Article
The value in a walkthrough like this one lives in the sequence: create a bot, trim its tools, set its model tier, define its gates, then let the team run. Writing that sequence down takes longer than recording it, because a video carries pacing and screen context that plain text has to reconstruct.
That gap is exactly what a tool like Skala Blog is built to close, and the same coordination idea applies to content: you already did the hard part when you recorded the video. If you have knowledge, explanations, interviews, or opinions sitting inside a YouTube video, you can paste the video URL into Skala Blog, let it transcribe the video, and generate a structured article from the transcript. The video does the talking, and the article carries the sequence forward for readers who will never watch it.
For a walkthrough this dense, the written version has one advantage the recording cannot match: you can re-read the tool table, the approval gates, and the timeout settings in any order, at your own pace, without scrubbing back through the timeline.
The hard part of a bot team is knowing where the line goes
The unlock in this article was never the bots, it was the coordination, and the same is true when you bring a finished video to a wider audience, since the knowledge is already there. Skala Blog
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits