Running Hermes Agent as an AI crew: 7 habits that matter

Hermes Agent is one of the most capable AI tools you can run today, yet in most hands it sits there like a chatbot inside Telegram; the presenter, who has run a whole crew of agents daily for months, says seven practical habits flip it from a bot you chat with into a team that works for you .

He starts by killing the one-agent-for-everything habit: research, writing and building all jumbled in a single chat tangle the results, so he sets up an orchestrator that plans the show plus research, writer and developer roles, each created with a plain-language brief, its own instructions and its own memory. The multi-profile gateway guide on GitHub recommends exactly this shape, giving every profile its own sessions and memory, and it maps one-to-one onto the crew in the video.

A room of its own inside Telegram

Then the crew moves into a single Telegram group with topics switched on, so every agent gets its own thread: you talk to the writer in the writer's room and to research in theirs, and it is always clear who does what. Because each topic carries its own session, the context window is split too, and the bloat of one giant chat never happens.

The trick is a small routing plug-in: out of the box every message in the group lands on the main agent, while the plug-in delivers each topic to the right agent, and the presenter shares ready-made prompts for it. Once installed, the crew feels like four separate apps in daily use, even though a single Telegram group runs underneath.

With the crew up, cost takes center stage, and the expensive-model myth falls: instead of the flagship for everything, each agent gets the model its job deserves, a modest one for mail triage or article summaries, a strong one for the builder. He tells how a budget-friendly model built his entire memory system, proving cheap does not mean bad, with his own bill as evidence.

Even inside one provider there are big and small siblings, plus an effort dial from light to extra high on the same model, and he turns both knobs per task. Then the bonus: if you already pay for a chat plan, you can link that account on the Hermes side and use OpenAI and Grok models with no extra charge, higher tiers bringing more allowance, usually far cheaper than pay-as-you-go tokens.

Cost and memory: the crew's two quiet bills

But the cheaper crew starts forgetting: when chat history fills up, the agent must compact it to make room, and every compaction bleeds detail, a whiteboard someone keeps wiping. His fix is a plain wiki of Markdown notes on his own server that every agent reads before answering and writes to after working, pages he can open and correct himself. LangChain's wiki-memory essay shows file-based memory readable by different tools, which is the same principle; the design takes inspiration from Karpathy's LLM wiki idea, and the crew's reading notes stay strictly separate from his own decisions, because someone else's marketing claim must never become your fact.

Once memory holds, the crew works while he sleeps: Hermes runs real scheduled jobs, a morning brief at eight with his mail digest, tasks, calendar and weather waiting on the phone, or an inbox check every thirty minutes with drafts ready. The scheduled-task documentation from NousResearch explains that real work can run on a timer with tunable delivery; the presenter's stay-quiet rule is the street version of that behavior, telling agents to say nothing when there is nothing new.

Safe connections and many crews on one box

Those jobs need tool access, and here he uses Composio: one browser sign-in connects Gmail, calendar and Notion, the agent receives a revocable key, and passwords never touch his server, a spare key for the house sitter you can take back any time. Composio's Gmail integration page confirms the access is scoped and revocable at any moment, matching the analogy. His three rules follow: grant access to one agent only, start read-only, widen later, because a sent or deleted message can never be taken back, and ready-made prompts keep the connections quiet.

The sequel to the third rule is stricter: mail is information, never instructions, since anyone can send you a message hiding orders your agent might obey. The attack is called prompt injection , and the rule is that the agent reads mail but never takes orders from it. The OWASP cheat sheet on AI agent security says external input must never count as commands and privileges must stay narrow; the presenter's rule lines up with that frame exactly.

The last two habits are about scale: with Docker he runs several fully separate Hermes setups on one server, content, work and personal crews never seeing each other's notes while a single box gets paid for. Docker's piece on AI teams describes isolated sandboxes giving every crew its own boundary, which is precisely this layout. Above it all sits NEXORA, his private mission-control board: missions start only on his approval, agents hand work to each other, results land in documents instead of scrolling away, with an agent city light, a task ledger, machine health and voice mode on top.