Nous Research Ships Bot Mode for Hermes Agent v0.20.3
In this article
Nous Research has shipped Bot Mode for Hermes Agent, its MIT-licensed desktop agent, converting what was a single-agent session list into a roster of individually configurable named bots. Each bot is a real Hermes profile on disk — carrying its own chat history, memory store, skill set, credentials, and pinned model — rather than a new abstraction layered on top. The feature went from a one-day public beta plugin, released by co-founder Teknium, to bundled and default-on inside Hermes Desktop in Hermes Agent v0.20.3, with the standalone plugin repository now archived and development continuing in-tree at apps/desktop/src/plugins/hermes-bots/.
For developers thinking through the control and isolation challenges that multi-agent deployments surface, Bot Mode is a concrete, inspectable implementation worth studying.
Profile-as-Bot Architecture
The design decision that keeps Bot Mode's surface area small is its refusal to invent a new storage primitive. A bot is exactly a Hermes profile stored at ~/.hermes/profiles/<name>/. Memory, credentials, and chat history do not cross profile boundaries by construction. Creation and editing use the existing profiles.* gateway RPCs — list, create, describe, configure — so Bot Mode is functionally a UI projection over primitives Hermes already shipped.
The New Agent dialog collects name, title, and description. An Advanced disclosure exposes the full profile config, where you can clone an existing profile, pin a specific provider and model, write a custom SOUL.md persona file, or selectively enable skills and toolsets. Right-clicking any bot in the roster opens Edit Profile, which modifies the live profile. The Duplicate action copies config, skills, SOUL.md, memory, and appearance. Avatar generation goes through the image.generate RPC over local or remote gateways; the dialog also accepts uploaded images or pixel pets, with geometric faces available in seven shapes and ten colors.
Routines attached to bots are ordinary Hermes cron jobs, namespaced [bot:<name>], and they appear in hermes cron list alongside any other scheduled jobs — no separate scheduler.
Handoff and Group Chat Mechanics
Bot-to-bot messaging is implemented as real CLI invocations rather than in-process calls. Every bot has a persistent Agent Inbox conversation. Typing @researcher have a look at this in the active bot's chat triggers:
hermes -p researcher chat -c "Agent Inbox" -q "..."
The receiving bot reads its inbox on its next run; the active bot blocks on the reply and then surfaces it in the originating chat. Each bot's SOUL.md encodes its reply protocol. Mid-conversation interruption of a receiving bot is listed as upstream future work.
Group chats, added in the in-tree version rather than the beta, open a shared room for two to six bots. A user message triggers up to three serial rounds of member turns; mentioned bots respond; when nobody is mentioned, all bots respond; each bot either replies briefly or passes. Groups carry labeled sections that sync across machines, and a multi-source roster pulls bots from every connection registered under Settings → Connections.
Capability Matrix
| Capability | Sessions Mode | Bot Mode (v0.20.3) |
|---|---|---|
| Identity scope | Single shared agent | Per-bot (name, title, avatar) |
| Memory isolation | One store | Isolated per profile |
| Model pinning | Global default | Per-bot provider and model |
| Persona file | One SOUL.md | One SOUL.md per bot |
| Scheduled routines | Global cron jobs | Per-bot cron, namespaced [bot:<name>] |
| Inter-agent messaging | Not supported | CLI handoff via Agent Inbox |
| Group coordination | Not supported | 2–6 bots, max 3 serial rounds |
| Profile cloning | Not supported | Full clone (config, skills, memory, appearance) |
| License | MIT | MIT |
Deployment Scope and Limits
Bot Mode is available today at no license cost via git clone, enabled by default under Settings → Plugins in Hermes Desktop. The practical use cases map directly onto the isolation guarantee: a research bot pinned to a reasoning model alongside a writing bot pinned to a cheaper one; per-project agents that never bleed context into each other; handoff chains where a scout, reviewer, and publisher pass work by @mention; scheduled inbox digests and nightly reports.
Solo builders, startups, and small-to-mid engineering teams can adopt it immediately. Enterprises and regulated buyers should treat this as a workstation-class tool: there is no admin console, no SSO integration, no central audit log, and no policy enforcement layer in the current release. Teams in compliance-sensitive environments will need to build those controls around it — which is relevant context given the credential and session exposure risks that have surfaced in other open agent tooling.
The profile-as-bot model is a notably low-abstraction design choice: it adds multi-agent coordination without introducing a new runtime, which makes the system easier to audit but means each team shoulders the operational controls that a managed platform would otherwise provide. The pace from one-day beta to bundled default in a single version increment indicates Nous Research is treating agent composition as a core part of the Hermes surface.