Hermes Bot Mode: Keep Specialist Agents in Their Own Lanes
Give Hermes specialists distinct profiles, scoped groups, human-led approvals, and clear limits for routines, model routing, and cross-machine messaging.
Start with one narrow specialist, not an autonomous department. By the end of this guide, you will have a small Hermes Bot that does one low-risk job while you remain responsible for every decision and result.
Build the first safe Bot
In current Hermes Desktop, Bot Mode is built in and on by default. Open the Bots tab in the left sidebar, then select New Agent in the roster. In Bot Mode, a Bot is simply a Hermes profile: the same underlying agent home, presented as a named specialist.
Create a fresh local Bot with a plain name, a short title, and a one-line description of its job. A useful first role is Researcher: it reads official documentation and produces a cited note. It does not publish, spend money, change production configuration, or contact customers.
Start with the smallest capability set
For a first Bot, the quick setup fields are enough. Advanced creation is where you choose fresh versus cloned creation, model and provider settings, a custom SOUL.md, and exact skill, toolset, and MCP-server access.
Do not clone a broadly privileged main profile just because it is convenient. Start fresh or explicitly limit its capabilities. Profile boundaries separate configuration, memory, skills, and chat history, but they are not an operating-system sandbox. On current Desktop builds, credentials also default to a shared OAuth or token pool with the main profile; treat that as a deliberate trade-off, not as isolation.
Give the Bot neither shell access nor customer-facing or production-writing authority on day one. Add a capability only after the Bot has a clear, reviewable need for it.
Give it one low-stakes task
Your first task should be reversible, easy to inspect, and able to produce a file or cited answer. For example: ask the Researcher to summarize one official documentation page and list its source links.
Use the Bot’s own canonical Bot Chat for that work. Inside this chat, /new and /reset are rerouted to /compact, so the relationship keeps one persistent history rather than forking into a scratch session. Ordinary sessions on the same profile still keep normal /new behavior.
Pause and check the result
Before creating a second Bot, confirm all four points:
- The Bot completed one low-stakes task while you watched it.
- Its enabled capabilities are limited to what that task needed.
- Its output can be checked independently against a source or a file.
- You understand which credentials it shares and which it does not.
If that is boring, it is working. Add another Bot only for a genuinely different job, such as a Critic who checks the Researcher’s citations. Keep shared directories intentional: one Bot should not overwrite another Bot’s working files by accident.
Use groups for review, with a human moderator
Groups can seat two to six Bots. A user message can trigger up to three serial rounds and ten messages total. That is a guardrail against a runaway conversation, not a throughput target.
Mention the Bot or Bots you want to hear from. Mentions scope the round; when nobody is mentioned, each member gets a chance to respond. Bots may pass when they have nothing useful to add, and each member keeps its own persistent Group: <name> session.
A good group task is: “Researcher, identify the documented behavior; Critic, check the source; show me the disagreement.” Read the return yourself and make the decision. Do not use group agreement as an approval mechanism, a deployment system, or proof that work is complete.
Treat Bot messages as a conversation, not a workflow
message_agent lets a canonical Bot Chat send an attributed message to another Bot’s canonical Bot Chat. It is useful for a bounded nudge, such as asking a Critic to inspect a file a Researcher just wrote. It is not a transactional hand-off: the sender receives an acknowledgement and a reply later as a background completion notification.
Retries are deliberately limited. A failed delivery is retried at most once for transient failures or context overflow; authentication, quota, and configuration failures do not auto-retry. Failed turns also carry a typed reason such as context_overflow, delivery_timeout, or provider_auth_or_access, which is more useful than guessing from provider prose.
For work that needs durable dependencies, an approval gate, a deadline, or a recovery record, use an explicit task system and a durable artifact store. Keep the actual file or record outside chat; chat is the discussion around the artifact, not the artifact itself.
Routines are Hermes cron jobs with visible ownership
A Bot Mode Routine is an ordinary Hermes cron job named [bot:<name>] …. Bot Mode makes the owner visible and puts each run’s result in that Bot’s chat history. It does not turn the schedule into a group discussion or add an approval step.
Use a Routine for one Bot doing one scheduled job. Know which machine owns the cron job, keep the job’s permissions narrow, and make healthy runs quiet while failures produce a useful alert. Do not create an unattended group to imitate a workflow engine.
Cross-machine Bot Mode has two routes
Cross-machine Bot Mode is supported, but it is not a substitute for durable orchestration. Choose the route deliberately.
Desktop relay works while a Hermes Desktop connected to both backends is running. Register the backends in Settings → Connections; the Desktop keeps the connections open, shows their Bots in the roster, and relays Bot messages between them. If the Desktop closes during a delivery, the sender is told that the reply did not arrive. The relay holds the connections and credentials; the gateways do not share each other’s authentication.
Peer messaging is the always-on option. It requires the remote machine to run the api_server gateway platform, a strong peer key, and private network reachability such as a LAN, Tailscale, or VPN. Configure it as infrastructure, not as a casual chat feature: use a unique peer name, protect the key in Hermes’ normal secret storage, test the failure path, and keep production artifacts independently verifiable.
Groups can also include Bots from connected machines. That is useful for a live human-led review, but it does not turn the group transcript into an audit trail. Store the evidence somewhere durable.
Verify model routing before you trust it
Bot Mode lets each Bot have its own model and provider selection. Do not use that label alone as a cost, policy, or compliance control. An upstream issue reported cases where a Bot Mode delivery used the default profile’s model instead of the recipient profile’s configured model.
Before a specialist role depends on a particular provider or cost tier, test the installed release with two lab profiles that have different model selections. Inspect the actual run evidence, not just the Desktop label. Keep that test free of real credentials and side effects.
When Bot Mode is the wrong tool
Use Bot Mode when you want a small, visible team of specialist profiles that you can supervise. Do not treat it as the control plane for a business or a production fleet.
Use a separate machine for real host-level isolation. Use an explicit scheduler and task system for unattended dependencies. Keep irreversible production, customer, financial, or access-control changes under direct human approval.
Done means
Your first Bot Mode setup is complete when you have one limited Bot that has performed a reviewable task under supervision, and you know its enabled capabilities and credential-sharing posture. If you add a Routine, you can identify the owning host and find its result in the Bot’s chat. If you need different model routes, you have tested those routes on your installed release.
What this article does NOT cover
This guide does not cover skill authoring, provider pricing and model selection, headless-server deployment, host hardening, or a durable multi-machine pipeline. Those are separate layers. Bot Mode is the human-facing cockpit over profiles, not the pipeline beneath it.
Related guides
- Hermes Agent Setup: First Run and the Provider Picker
- Hermes Agent Memory vs Skills
- Hermes Agent Tools: Enable Only What You Need
- Hermes Agent Cron Job Management
- Hermes Agent Gateway Troubleshooting
Research basis: Hermes Agent research dossier, Websites/content-strategy/research/hermes-agent/hermes-agent-research-2026-08-29.md (verified 2026-08-29).
Sources
Sources
Submit a take
Have a different read on this? Drop a comment below — your email isn't published, and I read every one. Nothing leaves the site until I approve it.