Nous Research just turned Hermes Agent from a single-assistant tool into a proper multi-agent workspace. Bot Mode, which shipped in v0.20.3, replaces the standard session list with a roster of named bots. Each one is a full Hermes profile with its own chat history, memory, skills, credentials, and pinned model. That means a research bot running a reasoning-heavy model can sit beside a writing bot on something cheaper, and neither leaks context into the other.

The architecture is refreshingly boring. A bot is just a profile living at ~/.hermes/profiles// with isolated config and storage. No new database, no exotic orchestration layer. The plugin rides existing RPCs — profiles.* for creation and editing, image.generate for avatars, and plain cron jobs namespaced [bot:] for routines. Bot-to-bot messaging works through a persistent Agent Inbox, where each bot’s SOUL.md teaches it how to respond. Typing @researcher have a look at this triggers a real CLI handoff: hermes -p chat -c “Agent Inbox” -q ”…”. Attribution arrives intact.

The rollout speed tells you something about how Nous operates. Co-founder Teknium ran a one-day public beta, collected bug reports, and folded the feature into the main desktop app. The standalone plugin repository is now archived; development continues in-tree. The bundled version added capabilities the beta lacked, including groups that sync across machines and group chats where two to six bots coordinate over up to three serial rounds. Mentioned bots respond, everyone responds when nobody is mentioned, and each bot can pass.

Who should actually use this? Solo builders, startups, and small engineering teams can adopt it today at no license cost. MIT licensing plus a default-on desktop plugin means zero procurement friction. But enterprises should read the fine print. There’s no admin console, no SSO, no central audit log, and no policy layer. That’s not a flaw — it’s a workstation tool, not managed infrastructure. Regulated buyers will need to build their own controls around it. The interesting question is whether this profile-based approach becomes the default pattern for open-source agent orchestration, or stays a desktop curiosity while heavier frameworks battle it out in the cloud.