Hermes Agent Bot Mode: Multi-Agent Teams, Explained

Bot Mode turns Hermes Desktop's session list into a roster of named Bots. Each Bot is not a new concept grafted on top of Hermes, it is an ordinary Hermes profile, stored at ~/.hermes/profiles/<name>/, that Bot Mode gives a face, a persistent chat, a schedule and a way to talk to its teammates. Nous Research's own documentation is explicit that the feature adds "no core patches, no background daemons, no extra storage" beyond that standard profile system, which is a meaningfully different engineering approach from, say, the real-time voice stack or the Agent-to-Agent protocol that shipped in the Herald Release three weeks earlier: those introduced genuinely new subsystems, where Bot Mode is largely a UI and system-prompt layer over primitives Hermes already had.

The feature has an unusually well-documented public history for an open-source plugin. It began as a standalone repository built by Nous Research co-founder Teknium, who ran it as a one-day public beta before it was pulled back for further work. It reappeared folded directly into the core desktop app via pull request #87886, "feat(desktop): bundle Bot Mode as built-in default-on plugin + core teammate protocol," which touched 52 files across 10 commits. That PR shipped inside Hermes Agent v0.20.3 (GitHub tag v2026.8.16.2), a patch release Nous Research published on 16 August 2026, and the standalone plugin repository was then archived, with development continuing in the main hermes-agent monorepo.

• Best for: Hermes users who want persistent, differentiated agent identities, a researcher, a coder, a scheduler, rather than one generic session that gets reconfigured for every task.
• Headline mechanics: named Bots with their own model, memory and skills; @mention handoffs; collaboration rooms of 2-6 Bots for up to 3 rounds per message; per-bot cron routines.
• Shipped: bundled default-on in Hermes Agent v0.20.3, 16 August 2026, via PR #87886; group-chat fixes followed in v0.20.4, 18 August 2026.
• Main caveat: it is a UI and coordination layer, not a new safety system. Every Bot carries the same real-system-access risk as any other Hermes agent, and live interruption of a mid-conversation Bot is explicitly listed as future work.

Bot Mode's path to shipping is unusually visible for a bundled desktop feature, because it started life as a public, MIT-licensed side project rather than an internal roadmap item. Teknium, a Nous Research co-founder and research engineer, first posted about it on X in early August: "Introducing Bot Mode for Hermes Agent. Bot Mode is an alternative to sessions mode, where you have one chat with each agent profile, or 'bot'. These bots can be given jobs, descriptions, profile pics, and communicate with your other bots! For one day we will do a public beta." That one-day framing was literal: the plugin, published as the standalone NousResearch/Hermes-Bot-Mode repository, was pulled back after its initial test window for further work before reappearing.

It reappeared with a second post from Teknium: "Reintroducing Bot Mode for Hermes Agent... These bots can be given jobs, descriptions, profile pics, and communicate with your other bots - They maintain their own memory." Nous Research's official account then announced the feature more broadly: "Introducing 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."

This sits on top of an already fast release cadence. Hermes Agent v0.19 "Quicksilver" shipped 20 July 2026 with an 80% cold-start speedup and smart approvals. Three weeks later, v0.20.0 "The Herald Release" shipped 3 August 2026 as the largest release in the project's history, adding real-time voice, the Agent-to-Agent (A2A v1.0) protocol, signed webhooks and desktop artifacts. Bot Mode then arrived through a rapid, lightly-documented patch train off the back of Herald: v0.20.1 (13 August), v0.20.2 (16 August), and v0.20.3 (also 16 August, tag v2026.8.16.2), the release that folded Bot Mode into the desktop app by default. Nous Research's own v0.20.3 notes are candid about the pace: they describe the release as consolidating roughly 250 commits across some 461 files and about 125 merged pull requests since v0.20.2, and state plainly that "full curated release notes for this window will ship with v0.21.0, which will document everything from v0.20.0 onward." In other words, Nous Research itself is flagging v0.20.1 through v0.20.4 as a fast, thinly-documented patch sequence rather than a fully written-up release, Bot Mode's bundling included.

For the fuller introduction to Hermes Agent as a product, see our original Hermes Agent explainer.

The core design decision behind Bot Mode is that a "Bot" is not a new object type. It is an existing Hermes profile: isolated configuration, memory, skills, credentials and chat history, all stored under ~/.hermes/profiles/<name>/, exactly as Hermes has always stored a profile. Bot Mode's job is to make a roster of those profiles feel like a team rather than a list of configuration folders. Nous Research's documentation for the original plugin describes it as "a pure UI layer over existing Hermes profile primitives," reading and writing through the standard profiles.* gateway RPCs (list, create, describe, configure) rather than introducing a parallel storage system, background daemon or core patch.

What PR #87886 changed was where that layer lives and how deeply it is wired into the agent's own behaviour. Rather than shipping as an optional plugin a user installs into ~/.hermes/desktop-plugins/, Bot Mode now lives at apps/desktop/src/plugins/hermes-bots/ inside the main hermes-agent monorepo and loads by default, with an off switch under Settings to Plugins. The PR also changes how bots learn to talk to each other. Previously, the "messaging other agents" protocol was appended to a user's own SOUL.md persona file; now it is injected directly into the system prompt as a stable-tier component, which means bot-to-bot handoffs work consistently across every session type, including headless CLI runs where there is no desktop UI reading a SOUL.md at all. A new bot_mode_protocol capability flag was added to the profiles list so a backend can explicitly signal whether it supports the protocol; newer builds gate the relevant SOUL protocol behaviour on that flag, while older gateways still work, they simply do not light up the newer coordination behaviour.

The PR's own validation numbers are worth noting because they are unusually specific for a UI feature: Nous Research reports 143 of 143 plugin-suite tests passing, 39 of 39 system-prompt tests passing, an end-to-end check for byte-stable prompt generation, and a measured cost of roughly 916 bytes added to the system prompt only when Bot Mode is enabled, with zero overhead on installations that leave it off. The roster itself is multi-source: the bundled plugin merges profiles from the active gateway with a "host agents" union roster, so a user with several registered Desktop connections sees bots from every one of them, with device-specific naming applied automatically when two connections have a profile with the same name.

The Bots pane and presence

Bot Mode replaces the flat session list with a left-side roster: one row per bot, showing its avatar, a preview of its latest message, and a timestamp. An "Active now" presence strip sits above the roster and lists every bot currently working, defined as the gateway-busy profile plus any bot that wrote within the last 90 seconds, without ever reordering the underlying roster, and it disappears entirely once the whole fleet is idle. Clicking a bot opens its chat directly; a context-menu "Sessions" option opens a filtered view of that bot's 200 most recent stored conversations.

Creating and configuring a bot

A quick "New Agent" dialog asks only for a name, a title and a description, enough to spin up a bot in seconds. An Advanced disclosure opens the full profile configuration underneath: clone an existing profile as a starting point, pin a specific provider and model (a bot can run on any model Hermes supports, and different bots in the same roster can run on entirely different models), write a custom SOUL.md persona file, or skip the bundled skill set entirely. Right-clicking an existing bot offers Edit Profile (change the avatar, title and description at any time, with the same Advanced section for live editing of per-skill and per-toolset enablement), Duplicate (a full clone of config, skills, SOUL.md, memory and appearance), and Delete Profile, gated behind the same destructive confirmation Hermes Desktop already uses elsewhere; the default profile cannot be deleted this way.

Groups and avatars

Bots can be organised into labelled groups via a right-click "Move to group" action, which sync alphabetically across every machine connected to the same gateway and disappear automatically once their last member leaves; ungrouped bots stay pinned above them. Identity is visual as well as functional: avatars can be procedurally generated geometric faces (seven shapes across ten colours, with blinking eyes that visibly "scan" while the bot is working), an uploaded image, an AI-generated portrait when an image backend is configured, or an animated pixel-pet companion that bounces beside the avatar during activity. Avatar and pet customisation is stored in plugin-local storage rather than the profile itself, so the underlying Hermes profile stays unmodified by cosmetic choices.

Routines and bot-to-bot messaging

Each bot can carry its own recurring tasks, backed by Hermes' existing cron system and namespaced [bot:<name>] <routine>, so "summarise my inbox every morning" sits attached to the bot that does it rather than to a global job list; runs still show up in both the Bots pane and the core hermes cron list output, and results land directly in that bot's own chat history. Bot-to-bot communication runs through a persistent per-bot "Bot Chat" conversation, delivered as a genuine CLI handoff (hermes -p <bot> chat --in ~ -c "Bot Chat" -Q -q "Message from 🤖 <sender> (@<sender>): ...") with the sender's identity attached to every message. Typing @researcher have a look at this in any chat hands the message to that bot, waits for its reply, and reports the result back, the same @mention mechanic that drives group collaboration rooms.

Collaboration rooms

The multi-agent centrepiece is the group chat, opened from any group's header, which supports between two and six bots in a shared room. A single user message triggers up to three serial rounds of member turns: any @mentioned bot responds first, and if nobody is mentioned every bot in the room takes a turn. Each bot can reply briefly or pass on a given round, and the room settles, stops taking further turns, once a full round passes with total silence. Two hard caps prevent runaway coordination loops: 10 messages per turn, and 3 rounds per user message. A bot can pull another bot into the conversation with @name, or escalate a genuine judgement call to the human by writing @user, which raises a visible "needs you" badge on the group's header. Each bot preserves its own persistent Group: <name> session underneath the shared room, so context survives exactly like any other Hermes conversation rather than being ephemeral to that one group chat.

Coverage of Bot Mode landed fast: both MarkTechPost and Open Source For You published detailed write-ups within a day or two of the v0.20.3 bundling, both confirming the same core facts, the Teknium-built origin, the default-on shipping in v0.20.3, the MIT licensing, and the two-to-six-bot, three-round collaboration room mechanics, independently of Nous Research's own documentation. That degree of independent confirmation across separate outlets is a reasonable signal that the headline claims about Bot Mode are accurate rather than optimistic marketing copy.

The more interesting reception, though, is the scepticism that showed up almost immediately in coverage aimed at practitioners rather than headline readers. One write-up put the concern plainly: "multi-agent setups have a habit of looking great in demos before you hit the coordination headaches in real use." The specific worries raised were concrete rather than generic: whether memory isolation between bots is airtight in practice, whether the coordination overhead of a six-bot room actually beats a single well-prompted agent for most tasks, how conflicting task completions from different bots in the same room get resolved, and the simple ongoing cost of maintaining separate persistent memory for every bot in a growing roster. None of these are flaws unique to Hermes; they are the standard open questions for any multi-agent framework, and Bot Mode does not yet have enough real-world runtime to have settled them.

Hands-on creator coverage was more mixed than the launch announcements. Alex Finn's video above frames Bot Mode as a direct answer to xAI's Grok Bot but is explicit about rough edges in his own testing, and this pattern, genuine enthusiasm about the concept paired with early-version friction in practice, tracks closely with how this site covered the Herald Release's own new surfaces (voice, A2A) three weeks earlier: a first version of something new, even from a mature project, is rarely as polished as the second or third.

Both Hermes Agent and Bot Mode are distributed under the permissive MIT licence, with the code now living directly inside the main NousResearch/hermes-agent repository rather than the archived standalone plugin. Nous Research is explicit that Bot Mode is architecturally minimal on purpose: no core patches, no background daemons, and no storage layer beyond the standard Hermes profile system. That is a genuinely good sign for auditability, there is comparatively little new attack surface to review, but it also means Bot Mode inherits Hermes Agent's existing safety posture wholesale rather than adding a stronger one. Every bot in a roster is a full Hermes agent with whatever system access that profile has been granted; multiplying the number of named, persistent agents on a machine multiplies the number of agents that need the same scoped-permissions discipline this site has flagged in every previous Hermes review, not a lighter one.

Two specific details from the plugin's own documentation are worth flagging directly rather than glossing over. First, bot-to-bot message delivery is per-invocation, not live: a receiving bot sees a handoff in its inbox the next time it runs, and the documentation lists "live interrupt of a mid-conversation bot" explicitly as upstream future work, meaning a bot mid-task cannot yet be reliably stopped by another bot's incoming message the way the Herald Release's barge-in voice feature can interrupt a spoken reply. Second, the graceful-degradation design (older gateways still show a working roster, just without Advanced editing or avatar generation) is a sensible compatibility choice, but it also means a user on an older gateway build may not get the full bot_mode_protocol system-prompt behaviour without realising it, worth checking directly if bot-to-bot handoffs seem unreliable.

Bot Mode adds no separate cost. Hermes Agent remains free and open-source under the MIT licence, and Bot Mode ships automatically as part of Hermes Desktop v0.20.3 and later, default-on, with no extra download, plugin install or subscription required. Existing users get it via hermes update; new users get it in any fresh install from that version onward. As with every prior Hermes release, the two real cost centres are unchanged: hosting (Nous Research continues to cite a roughly $5-a-month VPS as a workable baseline) and the API usage of whichever model each bot is pinned to. Bot Mode does introduce one practical cost wrinkle worth budgeting for: because each bot can be pinned to a different model and can run independently, a roster of several bots simultaneously active in a collaboration room can generate materially more parallel API calls than a single-session workflow did before, particularly if multiple bots are pinned to a frontier reasoning model rather than a cheaper option.

Nous Research had not published a standalone pricing page specific to Bot Mode at the time of writing, unsurprising given the feature carries no fee of its own; check the current Nous Portal plan list directly for the underlying model-access costs if using Nous Research's own hosted model access rather than a separate provider API key.

• No live interruption yet: bot-to-bot messages deliver on the receiving bot's next run, not instantly; a bot cannot yet be reliably interrupted mid-task by an incoming handoff.
• Coordination overhead is unproven at scale: the 2-6 bot, 3-round collaboration room design caps runaway loops, but early practitioner coverage flags real open questions about memory isolation, conflicting task reports, and whether multi-bot coordination reliably beats a single well-prompted agent.
• No new safety model: Bot Mode is a UI and coordination layer over existing Hermes profiles, not a new sandboxing or permissions system; every bot carries the same real-system-access risk as any single Hermes agent, multiplied by however many bots are active.
• Feature parity depends on gateway version: older gateways show a working roster but do not light up Advanced editing, avatar generation, or the full system-prompt-level teammate protocol without a newer build.
• Lightly documented patch train: Nous Research's own v0.20.3 notes acknowledge that full curated release notes for this window will not ship until v0.21.0, so some Bot Mode details are better documented in the plugin's own README than in the official changelog.
• More bots can mean more spend: a roster with several bots active in a room, especially pinned to different frontier models, can generate meaningfully more parallel API usage than a single session.

Bot Mode is best understood alongside two other things Hermes shipped in the same month, not in isolation. The Herald Release's A2A v1.0 protocol, covered in our full Herald review, solves a different problem: it lets a Hermes agent talk to agents built on entirely different stacks, over a shared wire protocol. Bot Mode solves the problem one level down, organising multiple agents that are already Hermes profiles, on the same install, into named identities with a roster UI, group chats and routines. A2A is about interoperability across vendors; Bot Mode is about identity and coordination within one Hermes install. The two are complementary rather than competing, and Nous Research's own delegate_task subagent-delegation mechanism (from Quicksilver) remains a third, more ad hoc option for one-off dispatched work that does not need a persistent named identity.

Against Anthropic's Claude Code, the closer comparison is cross-session messaging, which lets one Claude Code session discover and message another via ListAgents and SendMessage. That is a session-scoped, developer-facing tool for one agent to relay plain text to another running session; it has no roster UI, no avatars, no persistent named identities and no group-chat collaboration rooms. Bot Mode is a more consumer-facing, persona-driven take on a related idea, at the cost of being tied specifically to the Hermes Desktop app rather than working headlessly by default the way Claude Code's messaging tools do.

The most direct product comparison, and the one several creators reached for immediately, is xAI's Grok Bot, always-on AI teammates that each run on their own persistent cloud computer with a browser, filesystem and terminal, signing into a user's existing tools with their own credentials. The framing is genuinely similar, named, persistent, specialised agents rather than one generalist chat, but the execution model differs sharply: Grok Bot's teammates each get a dedicated cloud machine and operate software the way a person would, including inside tools with no clean API; Hermes' bots are lighter-weight profiles sharing one desktop app and one gateway, communicating through CLI handoffs rather than each getting isolated cloud infrastructure. Grok Bot is closer to giving each teammate its own computer; Bot Mode is closer to giving one computer several named colleagues who share its resources. Which is preferable depends heavily on whether the work in question genuinely needs isolated infrastructure per agent or simply needs a clearer division of labour inside one existing setup, and it is exactly the comparison Alex Finn's video above works through in detail.

Use Bot Mode if you already run Hermes Desktop and find yourself repeatedly reconfiguring one session for different jobs, a research pass one hour, a coding task the next. Building a specialist bot once, with its own model, skills and persona, and returning to it by name is a genuine quality-of-life improvement over rebuilding context every session, and the collaboration-room mechanic is worth trying for tasks that naturally split across a small number of distinct roles.

Be cautious if you need reliable live coordination between agents working on the same task in real time: per-invocation message delivery and the lack of mid-conversation interruption mean Bot Mode is better suited to handoffs and asynchronous collaboration than to tightly synchronised teamwork. And treat every bot in a roster with the same access discipline as a single Hermes agent; Bot Mode organises agents, it does not sandbox them.

Bot Mode is a smaller, quieter release than the Herald Release's voice stack or A2A protocol, and that is arguably its strength: it is largely a UI and system-prompt layer over primitives Hermes already had, which is exactly why Nous Research could take it from a co-founder's one-day public beta to a tested, default-on core feature in a matter of days. The public development history, a named individual's beta plugin, folded into the product with a specific, verifiable pull request, is unusually transparent for a bundled desktop feature, and the test numbers behind PR #87886 suggest real engineering care went into making the change close to free for anyone who doesn't want it.

Whether the collaboration-room mechanic holds up under real, sustained multi-agent workloads is genuinely still an open question, and early practitioner coverage is right to flag coordination overhead and memory isolation as things to watch rather than settled wins. What is not in question is the direction: Hermes Agent, like Grok Bot and Claude Code's own cross-session tools, is converging on the idea that one generalist chat session is no longer enough, and that persistent, named, specialised agents working together are where agent products are heading next.