Hermes Agent Update: Bots, Group Chats and a HUD That Drives Your PC

The latest Hermes desktop release turns a single assistant into a small staff of agents. Nous Research ships Bot Mode as a built-in part of the app, adds Discord-style group chats, gives every bot its own memory lane, and puts a HUD overlay on your screen that can read what you are doing and, on request, click through your software itself. Here is what the official release notes say, how the pieces work together in practice, and which trap quietly eats your token budget.

Most agents people rely on today belong to a single company. Claude Code sits inside Anthropic, Codex inside OpenAI. If that company stumbles, changes prices, or disappears, your daily work tool changes with it. Hermes Agent removes that dependency. The project from Nous Research is fully open source, works with any model, and switches between cloud APIs and local LLMs with one click. The loop stays the same no matter which model you pick: the model plans and generates, the agent's tools write files, read sources, run shell commands and drive the web, and the agent verifies results before taking the next step.

The other defining trait is self-improvement. After finished tasks, the agent writes Markdown skill files that capture how a recurring piece of work should be done next time. On the next run it builds on that skill instead of starting from zero. Around the skills sit a user memory, a fact and project memory, and a searchable database of past chats. The model weights never change; everything grows in the orchestration layer.

The headline change is Bot Mode, officially part of the desktop app since version 0.21.0. Every agent profile now has a name and a deterministic avatar face, and all bots live in a shared roster. On top comes a chat-app experience: group chats where several bots and you talk in one room, mentions to pull a specific bot into the conversation, and named, pictured rooms. Multi-agent setups used to mean plumbing; now they look like a messenger full of colleagues.

The practice demo in the video chains four bots around one topic: a research agent gathers the facts, a copywriter turns them into a report, a designer sets the visual and structural brief, and a coder implements the result. The bots hand work along the chain and return a finished package with research, text and design templates. A group-chat test adds a round of introductions where each bot presents itself in one sentence, and shows that you can address any bot directly or tag just one of them when you only need a social media post.

Every bot is configured separately: its own system prompt, its own personality, skills switched on or off per bot, and a model choice if you insist. Restricting a bot to the capabilities it actually uses also makes it leaner.

The most substantive part of the update is not the colorful roster, it is memory hygiene. Each bot keeps separate memories. Coding memories accumulate in the coding agent, copywriting memories in the copywriter, research facts with the research agent. Run everything through a single agent and its memory fills with noise that slowly degrades answer quality. Note that the context window itself is not the problem, because Hermes compresses it automatically. The bottleneck is what gets remembered, and Bot Mode keeps those lanes apart.

A concrete number from the video: in the shown instance the video-editing skill had already been used 252 times. The Capabilities view lists every skill with its usage count and lets you toggle them with a click. Work with the same agent for a few months and those counters show exactly where the real workload lives.

The automation side grew too: scheduled jobs create cronjobs, for example a daily AI news digest every morning at 9, and a Kanban board sorts tasks into triage, to-do and scheduled. Connecting Telegram now works through a QR-code quick setup taken straight from the dashboard, which landed officially in version 0.21.2. When your agent runs on a server, Telegram is the most direct control layer you can get.

The channel shares a hard-won lesson against the obvious optimization. Plenty of people recommend a model per role: a big one for the backend, another for frontend, a third for research. It sounds smart and performs badly once the agents actually talk to each other, for two reasons.

First, agents cross-check each other. Different models solve tasks differently, keep correcting one another, and burn tokens in loops that produce nothing. Second, prompt caching breaks. Repeated prompts to the same provider get cached and cost far less. Spread your bots across providers and the cache barely hits anymore, so token spend explodes.

The practical recommendation: run one model for every agent that interacts directly, and if you want variety, use different sizes of the same provider for orchestrator and supporting roles. Experimenting with other providers belongs in an isolated single agent, not inside your working staff.

The second update highlight is HUD mode, which has its own section in the official desktop docs. One click in the top right and a compact overlay sits on top of the current window. Hermes can see the screen and picks up context: while cutting a video in Shotcut, the channel asks the overlay what it is looking at, and Hermes immediately names the program and the task.

The overlay is not just an observer. Asked to take over the computer, Hermes clicks through the interface itself: the save dialog for an MP4 export gets handled, an empty audio track is added on track A1, and the agent reports back. That closes a real gap: software without an API or shell access can still be automated as long as a local gateway is allowed to drive the mouse. On a cloud agent the click would land on the server's own machine instead. Voice input works here too, listening runs through connected voice services rather than a plain dictation button.

The desktop app also ships its own browser. The video shows a typical task: open a channel's latest YouTube upload and search for a prompting paper in a second tab, then open both. Hermes navigates, clicks through YouTube, and sets up both pages. From there it can summarize the content, answer questions about what is on screen, or hand the browser back to you.

The best moment is self-improvement happening in the background: during that scene Hermes creates or improves a skill called Browser Research Handoff, capturing exactly the routine this session just used. The app also stores images and assets under Artifacts so they stay findable, and closing the browser returns you to the normal chat.

Local installs have one drawback: automations only run while your machine is awake, which fuels the current mini-PC boom. The video adds a cloud path through Abacus.AI, whose Supercomputer offering keeps agents running around the clock and bundles SSH, GitHub integration, databases and access to more than 100 models. The numbers shown in the video: the 20-dollar plan includes 30,000 credits, cheap flash and open-source models should stretch that into 15 to 70 million tokens, and an always-on server consumes roughly 8,000 to 31,000 credits per month according to the provider. The channel then connects the cloud agent to the desktop app over SSH, and one click in the interface flips the whole UI onto the cloud instance.

A side note from the same setup: the brand-new Union Alpha model, currently discussed on X for being cheap and benchmark-strong without any official announcement, got tested in the channel's robot benchmark. The verdict places it between GPT 5.6 Sol and Astra: better physics and rotation than Sol, below Astra and Fable. Where the model actually comes from remains unclear. That assessment is a video observation, not an official benchmark.

This release moves Hermes Agent from single assistant to small staff. Bot Mode turns multi-agent operation into a button, separate memories keep contexts clean, HUD mode reaches into software that used to be impossible to automate, and the cloud variant solves the always-on problem for roughly 20 dollars a month. The bundled lesson matters as much as the features: keep the model mix out of your chat room, give the staff one model, and quarantine experiments in isolated agents.