Ten Underused Hermes Features That Reveal Its Real Power
Persistent identity, session recovery, reusable skills, gateways, schedules, subagents, profiles, model switching, side questions, and event hooks.
Many people use Hermes as a slightly more capable ChatGPT: they connect Telegram, ask questions, and stop there. This misses the architecture that makes Hermes useful as agent infrastructure.
Its deeper features solve recurring operational problems. They preserve identity, recover earlier decisions, separate roles, schedule work, delegate tasks, and react to events. Together, they form an operating layer for ongoing work rather than a single conversation.
1. soul.md: persistent identity and boundaries
The soul.md file describes how the agent should behave. It can define voice, priorities, values, communication style, and approval rules.
The file should be specific but short. Statements such as “be helpful” add little. More useful guidance includes how to present uncertainty, when to ask permission, which language to use, and how to handle conflicting instructions.
Because the file is readable, identity changes can be reviewed and versioned. This is safer than allowing personality to drift invisibly through chat history.
2. Session search: recover earlier decisions
Long-running work produces decisions that are difficult to find manually. Session search allows the user to locate previous discussions, links, and conclusions.
Search is especially useful before repeating research or changing a plan. Ask for the earlier decision and the surrounding rationale, not merely a keyword match. Important findings should then be promoted into project notes or durable memory.
Session search is a recovery tool, not a replacement for organized documentation.
3. Skills: turn prompts into workflows
A skill captures a repeatable procedure. Instead of re-explaining how to prepare a report or process a lead, the user invokes the skill.
A good skill states its purpose, required inputs, steps, tools, output format, checks, and limits. It should be tested on several examples and revised when a real edge case appears.
Skills are one of the clearest ways Hermes becomes more useful over time. Improvement is stored in a visible artifact rather than left inside one forgotten conversation.
4. Telegram gateway: a genuine control surface
Telegram is more than a place to receive answers. Commands, topics, files, and notifications can make it a remote control surface for Hermes.
Different topics can map to separate projects. Scheduled reports can arrive in dedicated channels. Approval requests can be handled from a phone.
The gateway must restrict allowed users and protect the bot token. Final artifacts should be stored in the project system even if Telegram is the place where they are reviewed.
5. Cron jobs: recurring agent work
Cron jobs allow Hermes to operate on a schedule. Useful examples include morning briefings, weekly research, health checks, and project reviews.
The schedule should state time zone, model, data sources, output destination, and behavior when nothing changed. Trigger the job manually before trusting the schedule.
Recurring work should be reviewed periodically. A report that nobody reads should be paused, even if it runs successfully.
6. Subagents: parallel research and execution
Subagents can work on independent parts of a larger task. The main agent coordinates them and integrates their outputs.
Delegation works best when each subtask is narrow and the response format is defined. The main agent should not simply concatenate answers; it must resolve duplication and disagreement.
Subagents also need permission boundaries. A research agent rarely needs credentials that allow publishing or account changes.
7. Profiles: separate agent workspaces
Profiles allow one Hermes installation to support several roles. Each profile can use its own model, tools, memory, and identity files.
A client profile can remain separate from a personal profile. A coding profile can use a model optimized for software, while a research profile uses a model selected for long-context synthesis.
Profiles reduce context pollution, but they are not always equivalent to strong user isolation. Sensitive clients may require separate instances and credentials.
8. Model and provider switching
Different tasks justify different models. Hermes can switch providers or models without rebuilding the whole agent.
Use cheaper models for routine classification and premium models for difficult reasoning. Maintain an approved fallback and record which workflows have been tested with each model.
Changing the model can change tool-use behavior and writing style. Test the workflow after switching rather than assuming compatibility.
9. /btw: side questions without derailing the task
A side-question mode allows the user to ask something related without injecting it into the main working thread. This protects the task’s context and reduces the chance that a quick tangent changes the agent’s plan.
Use side questions for definitions, clarifications, and brief checks. Decisions that affect the main task should be restated in the main session so they become part of the authoritative context.
10. Hooks and BOOT.md: react to events and startup
Hooks run when defined events occur. A hook may respond to a new file, webhook, task-state change, or other signal. Hermes does not ship a built-in BOOT.md hook; the documented pattern is to create a ~/.hermes/BOOT.md checklist and a user-defined gateway:startup hook that reads it and runs a one-shot agent.
Startup actions should be lightweight and safe. Suitable examples include verifying required directories, loading a short status note, and reporting disabled integrations. Avoid heavy or destructive work that runs every time the service restarts.
Event hooks should validate their source and be designed to tolerate duplicate delivery. An external service may send the same webhook more than once.
Hermes as infrastructure
These features reveal the platform’s real shape. Identity provides behavioral continuity. Session search and memory preserve knowledge. Skills preserve procedures. Gateways provide access. Cron jobs and hooks provide initiative. Subagents and profiles provide delegation and isolation. Model switching provides flexibility.
The result is not simply a smarter conversation. It is a configurable infrastructure layer for agents, memory, workflows, messaging, and tools.
Begin with the feature that solves the most immediate problem. If old decisions are lost, use session search and project notes. If prompts repeat, create a skill. If work must happen regularly, add a cron job. If instructions conflict, split profiles. Purpose-driven adoption keeps the system useful instead of merely complicated.
Source video: 10 Hermes Agent Features Most People Aren’t Using, Sharbel A. Adapted and reorganized from the full transcript; calls to subscribe and other promotional passages were removed.