Source: raw/Hermes_Bot_Mode_Is_A_Cheat_Code._Here_s_How_I_Use_It.md (auto-caption transcript)

Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=p3_Ql5nq4t4 · Duration: 18:21 · Uploaded: 2026-09-22 · No sponsor named; setup and prompts offered as a free article in the description.

A hands-on walkthrough of Bot Mode in the Hermes desktop app, a week before Nous’s own announcement was recorded in this wiki (see Hermes Desktop). Sharbel explains that a bot is a profile, shows the settings that keep a team cheap, and then has his main agent build a three-bot Instagram content team that researches, drafts and cross-checks a carousel in a group chat.

Key Takeaways

  • A bot is a profile. Each has its own persona, model, tools and access, and memory. Bot Mode makes it easier to talk to them and lets them talk to each other. “It was never the bots, it’s the coordination.”
  • Three ways bots work together: one-on-one with you; direct, one bot messaging another; and a group chat with a cap on back-and-forth turns so bots cannot loop forever.
  • Creating a bot. New bot, then a name (used for @ tags) and a title (shown in the sidebar), then a one- or two-sentence job description. Hermes writes the bot’s soul.md (personality and rules) from that. You can clone from your default profile and pick a model. Tighten the soul afterwards; his sales bot got “never discuss prices unless I approve” and “always hand the result back to my main Hermes instance”.
  • Turn off tools each bot does not need. Skills load only when used, so they cost little. Tools and MCP servers load into every message, and a new bot inherits every tool and MCP your main bot has switched on. “Some tools have upwards of 50,000 tokens.” Give each bot one specialist connection (his sales bot alone gets the sales CRM).
  • Strong model for the orchestrator, cheap or local models for workers.
  • Memory. Only the orchestrator writes to a shared memory provider (Honcho, Mem0, Supermemory); other bots use built-in memory only, to avoid double saves.
  • Two settings he changes. Warm bots from 3 to 5, so switching is instant. Idle timeout from the default 10 minutes to 20 or 30, so long jobs do not quit halfway.
  • Gates in three layers, strongest first:
    1. Which tools a bot has. A bot without an email tool cannot send email.
    2. Hermes’s approval setting on smart: safe actions run, dangerous ones are blocked, the rest ask. Choices on a prompt: allow once, allow for the session, always allow, deny.
    3. A rule in soul.md. A soft layer the model tries to honour; never your only lock.
  • His three gates: spending money, sending something to a client, deleting things.
  • When not to build a team. A one-off job, or a single inbox, needs one good agent. Teams pay off on weekly recurring work.

The content-team demo

  • He asked his main agent (“Hermy”) to build an Instagram team: a researcher that monitors what works in the niche and a creator that turns it into posts, with Hermy as his only point of contact. The prompt ended “ask me as many questions as you would like until you reach clarity”.
  • Hermy asked onboarding questions, created both bots, wrote their souls (mission, reporting line, responsibilities) and created the group chat. He then gave the creator Instagram access, removed tools neither bot needed, and put them on a cheap subscription model.
  • In the group chat, @researcher, @creator or @everyone tags route a message. Hermy delegated research, the creator drafted a carousel, Hermy had the researcher check it for claims that were “off or not evidence-based”, and then reported back.
  • Glitches on camera: a researcher error, and a five-minute silence that turned out to be his GPT provider hitting its session limit.
  • Bots can be grouped by workflow (his “Daily Psych” group) to keep several teams organised.
  • After a job, tell the team to save what it learned so the next run needs less input: “you’re not just running a team. You are training one.”

Try It

  1. Build one bot first and learn its settings before building a team.
  2. For each bot, open capabilities and switch off every MCP and tool its one job does not use.
  3. Set the orchestrator to your strongest model and the workers to a cheap or local one.
  4. Write your own regret list (“what could a bot do that I’d actually regret?”) and enforce it at the tool level first, then with the smart approval setting.
  5. Ask your main agent to build the team for you, and to ask questions until it reaches clarity.

Open Questions

  • The default cap on group-chat back-and-forth turns is not stated.
  • Settings names (warm bots, idle timeout) come from the captions; confirm them in the desktop app.
  • This partly answers the open question in Hermes Desktop about bot-to-bot permissions: in this demo, permissions come from each bot’s own tool set and the approval setting, not from a separate inbox permission model.