Update — August 24, 2026: Hermes Desktop's documented preview rail now makes the chat a working surface, not just a transcript. It can render web pages, files, and tool output beside the conversation; the same app also exposes a project file browser, persistent terminal, and Git review pane. That combination matters more than a generic “preview” label: users can inspect what an agent is reading or changing without losing the chat context. Details and limits are covered below.
A Bot Mode demo of AI teammates splitting up a video-game project on their own is the moment Nous Research's multi-agent feature stopped being a slide and started looking like a workflow.
On August 18, 2026, at 12:39 AM, @NousResearch announced Bot Mode for Hermes Desktop: "Your agent profiles become a series of named Bots. Each Bot has its own role, model, memory, skills and profile picture; Bots can use any model and even communicate with each other. Build a specialist Bot once to use it forever." The post had crossed 924.3K views at time of writing and was reposted by Robert Scoble — real traction for a feature ship, not just a routine changelog line.
The most convincing part isn't the announcement copy, it's the demo. Nous co-founder Teknium shared a video from @tonbistudio showing Bot Mode's group chats in action, captioning it "Here's @tonbistudio's video on Bots Mode in Hermes Agent!" Tonbistudio's own description: "Watching a team of bots split up a large game project based entirely on their own experience and specialties with little input from me is pretty crazy (and just a bit creepy) to watch." One of the Bots in the clip runs on Qwen. Nous Research's reply: "Banger as usual."
That's the pitch for readers of explainx.ai's coverage on multi-agent orchestration and agent harnesses: a consumer desktop app just shipped named, model-agnostic, inter-communicating agents as a default UI primitive, not a research preview behind a waitlist.
Hermes Desktop demo footage, hosted by Nous Research.
TL;DR
| Question | Answer |
|---|---|
| What shipped? | Bot Mode — named, persistent agent profiles ("Bots") in Hermes Desktop |
| When? | Announced August 18, 2026, 12:39 AM via @NousResearch |
| What's new vs. a normal agent profile? | Each Bot has its own model, memory, skills, and avatar — and Bots can message each other |
| Where do I get it? | hermes-agent.nousresearch.com — macOS/Linux installer script, or native installers |
| License | MIT, per the Hermes Agent product page |
| Current version | v0.20.3, per the product page (as of this post) |
| Cost | Free tier exists; Plus/Super/Ultra add monthly model credits via a Nous Portal |
| Best demo so far | @tonbistudio's group-chat video: bots dividing up a game-dev project by specialty |
What Bot Mode actually changes
Before Bot Mode, an "agent profile" in Hermes Desktop was a single configuration — one model, one memory store, one skill set, attached to one conversation thread. Bot Mode reframes that profile as a named, reusable identity. Per Nous Research's own description, each Bot carries:
- A role — what the Bot is for (e.g. "level designer," "QA," "narrative writer")
- Its own model — any model Hermes Desktop supports, chosen per-Bot, not globally
- Its own memory — a Bot's learned context doesn't bleed into another Bot's
- Its own skills — the skill set a given specialist has picked up
- A profile picture — a visual identity inside multi-Bot chats
The reusability point matters as much as the multi-agent point: "Build a specialist Bot once to use it forever." That's a direct answer to the throwaway-agent problem in most chat-based tools, where useful configuration (system prompt, tool access, learned context) gets rebuilt from scratch every session. It's the same instinct behind persistent memory in Claude.md files and Memory.md patterns for coding agents — except here it's packaged as a consumer-facing, named character rather than a config file a developer edits by hand.
The feature that makes this "multi-agent" rather than just "agent profile management" is the second half of the announcement: "Bots can use any model and even communicate with each other." That's the mechanic tonbistudio's video demonstrates — a Qwen Bot and its teammates dividing a game-dev backlog in a shared chat, deciding among themselves who handles what, with the human mostly watching.
Where Bot Mode fits inside Hermes Desktop
Bot Mode isn't a standalone product — it's a new mode inside the existing Hermes Desktop app, which explainx.ai has covered since its OpenRouter #1 ranking run and its expansion into a self-hosted web UI. Per the current product page at hermes-agent.nousresearch.com (tagline: "The Agent That Grows With You"), Hermes Desktop's six core features are:
| Feature | What it does |
|---|---|
| Connect ("Lives Everywhere") | Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI — one agent, one memory, every surface |
| Remember ("Persistent Memory") | Learns your projects, auto-generates skills, keeps solved-problem history |
| Schedule ("Focused Automation") | Natural-language scheduling for reports, backups, briefings, run unattended via a gateway |
| Delegate ("Tasks Multiplied") | Isolated subagents with their own conversations, terminals, and Python RPC scripts |
| Search ("Browse the Web") | Web search, browser automation, vision, image generation, text-to-speech, multi-model reasoning |
| Experiment ("Isolated Sandboxing") | Five backends — local, Docker, SSH, Singularity, Modal — with container hardening and namespace isolation |
Bot Mode sits closest to Delegate, but it's a step up in kind, not just degree. Delegate's isolated subagents run one task and disappear; Bot Mode's Bots are named, persistent, and reusable across sessions — closer to how types of AI agents discussions describe specialist roles than to a one-shot subagent call. Reading the launch against the full feature list also answers the obvious follow-up question — "is this the whole product?" No: Bot Mode is one new UI layer over an app that already does multi-platform messaging, memory, scheduling, and sandboxed execution. Readers evaluating Hermes Desktop for the first time should weigh the whole stack, not just the viral clip.
Hermes Desktop is downloadable today: native builds for macOS, Windows, and Linux, with a one-line installer for macOS/Linux —
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
The product page lists the current build as v0.20.3, MIT-licensed, with Free, Plus, Super, and Ultra tiers — paid tiers bundle monthly model credits and access to "300+ cutting-edge models and built-in tool use" through a Nous Portal.
What the interactive preview pane actually changes
The current Hermes Desktop documentation describes a right-hand preview rail that renders web pages, files, and tool outputs while chat continues in the center. It sits inside a broader inspection workflow:
| Surface | What it adds |
|---|---|
| Preview rail | Web pages, files, and tool results stay visible beside the conversation |
| File browser | Follow the agent through the active project's directory and preview files in place |
| Persistent terminal | Run or watch commands without leaving the app; selected output can be added back to chat |
| Git review pane | Inspect last-turn or branch diffs, stage changes, commit, push, or create a PR |
That does not mean every rendered page is automatically safe for autonomous interaction, nor that a preview turns Hermes into a general browser replacement. It means the desktop harness closes an important feedback loop: the agent acts, the artifact remains visible, and the human can inspect or intervene without mentally reconstructing tool output from a text summary. The same control principle appears in visual agent review workflows: keeping the real artifact beside the conversation reduces translation errors between “what the agent says it changed” and what actually changed.
What people are asking
"Is this actually multi-agent, or just multiple chat tabs?"
The distinguishing claim is inter-Bot communication — Bots message each other, not just the human. That's the difference between running five separate agent windows and running a group chat where specialists coordinate on their own. Tonbistudio's demo backs this up: the bots split the game-dev backlog among themselves "with little input" from the human operator, rather than each waiting on separate instructions.
"Can I mix models across Bots — Qwen for one, Claude or GPT for another?"
Yes, per the announcement: "Bots can use any model." The demo video specifically calls out a Qwen Bot among the team, and Hermes Agent's broader product page already advertises access to 300+ models via the Nous Portal on paid tiers. That model-agnosticism is consistent with how Hermes Agent compares to OpenClaw — both are built to be stack-agnostic rather than locked to one vendor's model.
"How does this compare to Anthropic's multiagent orchestration?"
Different surface, similar idea. Anthropic's multiagent orchestration for Claude Managed Agents coordinates up to 20 specialists inside a managed cloud environment, aimed at enterprise workflows. Bot Mode is a desktop-app feature: you name each Bot, assign its model and memory yourself, and watch the group chat happen locally on your own machine. It's closer in spirit to how YC's QM harness frames "multiplayer" agents, except QM targets whole-company Slack/web deployments while Bot Mode targets a single user's desktop.
"Is this production-ready, or a launch-day demo?"
It's day one. The tweet and demo video establish what Bot Mode does; they don't establish how it behaves at scale — how many Bots one chat can hold before coordination degrades, how memory isolation holds up across long-running group projects, or how tool/skill conflicts between Bots get resolved. Treat the game-dev demo as a proof of concept worth trying, not a benchmark of guaranteed reliability.
"Do I need to already run Hermes Desktop to use Bot Mode?"
Yes — Bot Mode is a feature inside the existing app, not a separate download. If you've never used Hermes Desktop, start with the installer above and the Hermes Agent walkthrough for VPS/Telegram/CLI setups to understand the base app before layering Bots on top.
Honest limitations
- No public benchmark or reliability data on Bot Mode yet — everything currently public is the announcement tweet, one third-party demo video, and the product page's general feature claims.
- Coordination overhead is unproven at scale. A two-or-three-Bot game-dev demo doesn't tell you how a ten-Bot team handles conflicting decisions or shared file state.
- Model costs still apply per Bot. Free-tier users are still bound by whatever the underlying model(s) charge if they're not using Nous Portal credits; running several Bots on separate paid models multiplies spend.
- This is Nous Research's own demo ecosystem — Teknium sharing tonbistudio's video is an authentic community reaction, but it's still a friendly amplification, not independent third-party testing.
Related on explainx.ai
- Hermes Agent hits #1 on OpenRouter
- Hermes Agent vs OpenClaw comparison
- Top 10 things you can do with Hermes Agent
- Hermes WebUI — self-hosted agent interface
- Hermes Agent remote VPS + Telegram + CLI guide
- Anthropic's Dreaming + multiagent orchestration for Claude Managed Agents
- YC open-sources QM — company-wide multi-agent harness
- What is an agent harness? Complete guide
- Types of AI agents — complete guide
Version numbers, pricing tiers, and product-page details reflect the Hermes Agent site (hermes-agent.nousresearch.com) and Nous Research's X posts as of August 18, 2026, and may change as Bot Mode matures.