7 Hermes Agent Upgrades You NEED To Implement Today

Oct 6, 2026

Read: ~7m · You save: 3 min

7 Hermes Agent Upgrades You NEED To Implement Today

Turn Hermes from a chatbot into a working AI agent crew with seven practical upgrades: teams, cheaper models, memory, cron jobs, safe tools, and mission control.

Hermes Agent is among the most capable AI tools available. For most users, however, it functions as a chatbot inside Telegram. The author has run Hermes daily for several months using a group of agents rather than a single one. The following seven changes turn Hermes from a chat interface into a coordinated team.

1. Build a Crew of Specialized Agents

Using a single agent for research, writing, and coding causes context to mix over time. Splitting the work into defined roles keeps each agent focused.

Orchestrator: the main agent that plans and coordinates the work.
Research: finds and summarizes information.
Writer: produces written content.
Developer: builds code.

A terminal is not required. Ask the main agent to create a persistent agent, for example: "Create a new persistent agent called research. Its only job is to find things out and summarize what it finds. Give it its own instructions and its own memory. Tell me when it's ready." The main agent builds the agent. Agents can also be created one at a time or several at once.

Assigning each agent a Telegram topic

Create one Telegram group.
Enable topics in the group settings.
Assign each agent its own topic.

Each topic functions as a separate session. Each agent keeps its own conversation history, and the user always knows which agent is working on which task. Separate sessions also prevent one context window from filling up, which occurs when all agents share a single chat.

Limitation and workaround: By default, Hermes sends every message in the group to the main agent. A small plugin routes each topic to its designated agent.

2. Match Each Agent to a Model That Fits Its Task

Each agent can run its own model. Choosing a model by task reduces cost without reducing output quality where the task is simple.

Complex tasks: A developer agent that writes code benefits from a stronger model.
Routine tasks: An agent that checks email or summarizes articles does not require the most expensive model.
Model families: A single provider offers multiple tiers. OpenAI offers larger models, such as GPT-Xtra and GPT-5.6, and smaller models, such as GPT-5.5 and GPT-5.4 mini.
Effort levels: Many models allow effort settings from light to extra high. The effort level should match the requirements of each task.

The author reports that a complete memory system for AI agents was built with MiniMax M3, a model that cost substantially less than the leading models from Anthropic and OpenAI.

Lower-cost access options

ChatGPT plan: Users with a ChatGPT subscription can sign in to Hermes with that account and use OpenAI models at no additional cost.
Grok subscription: A SuperGrok or X subscription provides Grok access. Usage under a subscription is generally less expensive than pay-as-you-go API usage. Higher plan tiers provide higher usage limits.

3. Give Agents Persistent Memory

Agents do not have long-term memory. When chat history reaches its limit, the agent summarizes it to free space. This process, called compaction, removes details. Instructions given early in a conversation can disappear after several long sessions.

Solution: a shared wiki

Store notes as plain Markdown files on a personal server.
Each agent reads the wiki before it responds.
Each agent writes to the wiki after completing work.
The user can open any page, review what the agents have documented, and correct inconsistencies directly.
A dedicated librarian agent can maintain the wiki automatically.

The wiki is accessible to any AI agent that reads it, including Claude Code and Codex. Those tools can read the same wiki and resume work with the same context.

Separation rule: Keep the user's own facts, such as decisions and client information, separate from material the agents retrieve online. Information from external sources should not become a stored fact without verification.

The approach is based on Andrej Karpathy's LLM wiki concept.

4. Schedule Recurring Tasks With Cron Jobs

Hermes can run tasks on a schedule through cron jobs. Scheduled jobs perform work, not only reminders.

Examples

Morning brief: At 8:00 a.m., an agent sends a summary that includes unread emails, tasks requiring action, the calendar, and the weather for the day.
Email monitoring: An agent checks email every 30 minutes and drafts replies where appropriate.

Avoiding unnecessary notifications: Most checks find no new items. Each scheduled job should instruct the agent to remain silent when there is nothing new to report. Hermes includes a built-in method for this, so no message is sent.

5. Connect Tools Securely

Agents require access to external services such as Gmail, Google Calendar, and Notion. The author uses Composio to provide these connections.

How the connection works

Sign in to Composio once in a browser.
Connect the tools the agent should use.
Connect the Hermes agent to Composio.
The agent receives a key that grants access to the connected tools.

Composio offers a free plan that provides substantial functionality. Access can be revoked at any time. Passwords for the connected services never pass through the user's server. The arrangement resembles giving a house sitter a spare key that can be withdrawn at any time, without giving them the original key.

Three access rules

Assign access to one agent only. A developer agent does not need access to the inbox and should not have it.
Begin with read-only access. A finance dashboard, which the author plans to cover in a forthcoming tutorial, reads receipts from Gmail. It cannot send or delete emails. Additional permissions can be granted later. A sent or deleted email cannot be recovered through permission changes.
Treat email as information, not instructions. Anyone can send an email containing hidden instructions intended to make an agent act. This technique is called prompt injection. Agents should read emails as content and should not follow commands contained in them.

6. Run Multiple Isolated Setups on One Server

Separate crews can be created for content creation, work, personal use, or individual clients without purchasing additional servers.

Docker allows one server to run several independent Hermes setups.
Each setup has its own isolated workspace, including its own agents, memory, and Telegram bot.
No crew can access the notes of another crew.
Costs are limited to a single server.

7. Monitor Agents From a Private Mission Control Dashboard

As the number of agents and setups grows, managing them across Telegram topics and the terminal becomes difficult. The author built a single dashboard, named Nexura, to manage all agents.

Core functions

Mission approval: Users launch a mission and review the execution plan. Nothing runs until the user approves it.
Work tracking: Results are delivered to documents, where they can be retrieved later, rather than through chat history.
Direct chat: Users can chat with any agent from the dashboard.
Activity view: An "agent city" display shows which agents are currently working.
Task log: Each task is recorded in a database with a summary of the task, the time performed, the AI model used, and the agent's name.
System monitoring: The dashboard shows machine health and all scheduled jobs, including the 8:00 a.m. morning brief.
Voice mode: Users can speak to agents, and agents can respond by voice.

Security: The dashboard runs on the user's own machine. Only devices owned by the user can access it. The author identifies security as the highest priority, because agents hold information about the user's email, clients, and other sensitive data.

Introduction

Hermes is presented as more than a chatbot. The speaker has run it daily for months as a crew of agents, and the video promises seven tips, with tip seven called the favorite.

Hermes is described as one of the most powerful AI tools available right now.
Most people use it as a chatbot inside Telegram.
The speaker has run Hermes every day for months, as a crew of agents rather than one agent.
The video is not sponsored; the tools mentioned are ones the speaker genuinely uses.

Tip 1: Build a crew, not one agent

Instead of one agent handling everything, split work among specialized agents, each with a defined job and its own memory and topic in a Telegram group.

Use a main agent called Orchestrator to plan and run the show.
Example specialized agents are research (finds things out), writer (writes), and developer (builds).
Agents can be created by asking the main agent in chat, for example: create a new persistent agent called research whose only job is to find things out and summarize them.
Create one Telegram group and turn on topics in the group settings to give every agent its own topic.
Each topic is its own session, so agents keep separate conversations and the context window does not blow up.
Out of the box, Hermes sends every group message to the main agent; a small plugin routes each topic to the right agent. The prompt is in the video description.

Tip 2: You don't need the most expensive model

Match each agent's model and effort level to its task. Cheaper models are often sufficient, and ChatGPT or Grok subscriptions can provide model access without extra API cost.

A whole memory system for AI agents was built by MiniMax M3, which cost much less than top models from Anthropic and OpenAI.
Every agent can run its own model; a coding agent may need a stronger model, while email checking or article summarizing does not.
Within one provider there are model families, such as larger GPT-5.6 and cheaper GPT-5.5 or GPT-5.4 mini.
Effort levels can be adjusted for the same model, from light up to extra high.
Signing in to Hermes with a ChatGPT plan allows use of OpenAI models without extra payment.
A super Grok or X subscription also works; higher plans give more usage, which is typically much cheaper than pay-as-you-go API pricing.

Tip 3: Give your agents a real memory

Agents lack long-term memory and lose details when chat history is compacted. The fix is a shared wiki of plain markdown files on your own server that agents read before answering and write to after working.

Chat history fills up, and the agent summarizes it (compaction), which causes details to be lost.
Example: instructions on how reports should be written given on Monday are probably summarized away by Friday.
The speaker gives the crew a wiki of plain notes in plain markdown files on their own server.
Every agent resets before answering and writes to the wiki after working.
Markdown files can be opened, read, and manually corrected for inconsistencies.
A librarian agent can maintain the wiki automatically.
Any AI agent, such as Claude Code or Codex, can read the same wiki and get up to speed quickly.
Rule: keep your own facts (decisions, clients) separate from articles and content agents read online, since someone else's marketing claim should never become your fact.
The concept is inspired by Andrej Karpathy's LLM wiki concept.

Tip 4: Let your agents work without you

Hermes can run real scheduled work through cron jobs. Recurring checks should stay silent when nothing is new, so the phone is not buzzing needlessly.

A cron job is used for scheduled real work, not just reminders.
Example: every morning at 8:00 a.m. the agent sends a morning brief with a summary of unread emails, tasks, calendar, and the day's weather.
Example: an agent can check email every 30 minutes and draft replies.
Trick: in every scheduled job, tell agents to stay quiet if there is nothing new to report; Hermes has a built-in way to send nothing at all.
The exact prompt for this is in the video description.
Agents need tools to access apps like email, which leads to tip five.

Tip 5: Connect your tools the safe way

Connect agents to apps like Gmail, calendar, and Notion through Composio, using least-privilege access, read-only starts, and treating email content as information rather than instructions.

Composio is used to connect Gmail, calendar, Notion, and other apps; it has a large pool of tools and a free plan to start.
You sign in once in the browser and connect the tools you want; the Hermes agent then gets a key to access them.
Access can be cancelled at any time.
Password access to these tools never touches your server, analogous to giving a house sitter a spare key that can be taken back.
Rule 1: give access to one agent only, not the whole crew; a developer agent probably has no reason to read your inbox.
Rule 2: start read-only. A finance dashboard reads receipts from Gmail and can never send or delete email. More access can be given later, but sent or deleted email cannot be recalled.
Rule 3: tell agents that email is information, never instructions, to defend against prompt injection, where hidden instructions in an email try to make agents leak information.

Tip 6: Run several Hermes setups on one server

Use Docker to run several completely separate Hermes setups on a single server, each isolated with its own agents, memory, and Telegram bot.

Separate crews can be made for content creation, work, personal use, or each client.
With Docker, one server can run several fully separate Hermes setups.
Each setup has its own isolated workspace: its own agents, memory, and Telegram bot.
No crew sees another crew's notes, and only one server must be paid for.

Tip 7: Build your own mission control

Build a single private mission control screen to launch missions, approve plans, monitor agents, check machine health, and view scheduled jobs, with security as the highest priority.

The speaker's mission control is called Nexura.
Missions are launched and a plan is set; nothing runs until the user approves it.
Agents pass work between themselves, and results land in docs for later review instead of scrolling through chats.
You can chat with any agent from the dashboard.
An agent city lights up when an agent is working, showing who is busy at a glance.
An agent log database records each task summary, the time performed, the AI model used, and the agent's name.
The dashboard shows machine health and every scheduled job, including the 8:00 a.m. brief.
Voice mode lets you talk to agents, and they can talk back.
The dashboard runs on your own machine and only your owned devices can reach it, because the crew knows your emails and clients.