AI · AIC-111

Self-Improving AI Agent with Hermes

Self-Improving AI Agent with Hermes is an advanced hands-on course in running Hermes Agent on a server around the clock, controlling it from Telegram, securing it with approvals and allow-lists, and building skills that improve from feedback, scheduled jobs and subagents. It suits developers, IT teams and existing Hermes users, and you leave with an agent that runs scheduled work and reports through a chat app.

Updated 13 Nov 2026

From 4,410 THB / person 4,900 −10% excl. VAT 7% · group rates available

Corporate training quote Join the waiting list

PDFDownload the course outline

Duration12 hours · 2 days
FormatOnsite / live online
Next roundOn request
CertificateIncluded

Course overview

Once you are comfortable using an AI agent on your own computer, the next step is an agent that works for you all the time: no laptop left switched on, instructions sent from your phone wherever you are, scheduled work that reports back on its own, and skills that get better from the work it repeats every day. An agent that lives on a server and can run commands by itself also brings higher risk, from who is allowed to give it instructions to file permissions, secret tokens and personal data, so it has to be designed to be both capable and safe from the start.

This course follows on from Local AI Agent with Hermes and Ollama. You move Hermes Agent from your own machine to an always-on setup on a VPS, in Docker or under WSL2, control it from Telegram, configure approvals, allow-lists and writable folders, and onboard the agent with identity, user and memory files. Day 2 focuses on skills that improve from feedback, scheduled jobs with cron, connecting MCP to a database, plugins and hooks, parallel subagents, a knowledge base built from notes and handing coding tasks to coding agents, and ends with an agent that runs scheduled work and reports through a chat app. (2 days, 6 hours per day, 12 hours in total, Advanced level.)

What you’ll gain

Install Hermes Agent on a VPS, in Docker or under WSL2 and run it as an always-on service
Connect Telegram to give instructions, approve commands and receive reports from your phone
Configure approvals, an allow-list of users, writable folders and secret storage safely
Onboard the agent with SOUL.md, USER.md, memory and context files, and manage context to save tokens
Build skills that improve from feedback and keep the skill library tidy
Schedule jobs with cron and deliver the results to a chat app automatically
Connect MCP to a database, install plugins and use hooks to log activity
Delegate work to parallel subagents, use notes as a knowledge base and hand coding tasks to coding agents

Who this course is for

People who have completed Local AI Agent with Hermes and Ollama or already use Hermes Agent
Developers and IT teams who want a personal or team AI agent that runs all the time on a server
System administrators and DevOps staff who want an agent to check systems, summarise reports and send scheduled alerts
Team leads and automation builders who want to move from workflows to an agent that learns from its work

Prerequisites

Completed Local AI Agent with Hermes and Ollama, or have already installed and used Hermes Agent with skills and MCP
Comfortable in a terminal, with basic Linux commands, SSH and an introductory knowledge of Docker
Your own laptop, plus your own Linux VPS, or Docker / WSL2 on a machine where you can install software
A Telegram account for creating a bot, and an API key from a model provider or a local model endpoint

Curriculum

Course Details

This course follows on from Local AI Agent with Hermes and Ollama (AIC-109). It shows you how to move Hermes Agent (open source from Nous Research, MIT licence) onto a server as an always-on agent controlled from Telegram, with skills that improve from feedback, cron jobs, MCP, plugins, hooks and subagents. It runs for 2 days, 6 hours per day (12 hours in total, 09:00-16:00), Advanced level. Basics such as local installation, Ollama and building a first skill are covered

in AIC-109 and are only briefly recapped here. Learners bring their own VPS or run Docker on their own machine, plus their own Telegram bot account and an API key or local model; the institute does not provide servers, accounts or credits. The course differs from Personal AI Agents with OpenClaw and Codex, which uses OpenClaw and Codex, and from Local LLM Deployment with Ollama and vLLM, which focuses on deploying and serving models that can act as an endpoint for the agent built here.

Day 1 An Always-on Agent on Your Own Server

Section 1: Hermes Architecture for an Always-on Agent

A short recap of AIC-109, then the step up to an agent that runs on a server and takes work from a chat app
The parts of Hermes: the agent loop, tools, memory, skills, the gateway and terminal backends
Choose a model provider, switch models with /model and plan a main and a fallback model by task and budget
Sessions, the web dashboard and the hermes doctor command for health checks
Lab: draw the plan for your own agent: where it runs, who can instruct it and what it can reach

Section 2: Lab: Installing on a VPS, in Docker or under WSL2

Install on a Linux VPS with the official script, a separate user and SSH key login
Run the official Docker image with data kept on a volume, so upgrades never lose your settings
Windows users: install inside WSL2, enable systemd and have the gateway start with the machine
Install the gateway as a service, update versions and manage server files with VS Code Remote SSH
Lab: a Hermes gateway running on the server that comes back by itself after a restart

Section 3: Lab: Controlling the Agent from Telegram

Create a bot with BotFather and keep the token in a .env file with restricted read permissions
Configure it with hermes gateway setup and set the user IDs allowed to give instructions
Chat commands: /new, /status, /stop, /model and /sethome for the reporting channel
Send files and images to the agent, and try voice messages as an optional demo
Lab: instruct the agent from your phone to check the server and reply with a short summary

Section 4: Workshop: Securing an Agent That Is Always On

Real threats: leaked tokens, strangers messaging the bot, prompt injection from web pages or files, and destructive commands
The smart and manual approval modes, a deny list, and why YOLO mode does not belong on a server
Allow-lists and DM pairing, restricting writable folders, and the Docker backend as a sandbox
Keeping secrets out of the sandbox, firewall and SSH on the VPS, activity logs, and PDPA when personal data is involved
Workshop: work through a hardening checklist and prove that risky commands and unauthorised users are really blocked

Section 5: Workshop: Onboarding the Agent with Identity, User and Memory Files

SOUL.md sets personality and tone, and /personality adjusts it temporarily for a session
USER.md and MEMORY.md: what they hold, their size limits and how the agent keeps them organised
Project context files such as AGENTS.md, and searching past sessions instead of stuffing everything into the prompt
Save tokens with /compress, fresh sessions and the right model for each task
Workshop: onboard the agent with the user, the work and the team's rules, then test it through Telegram

Day 2 Self-Improving Skills and Multi-Agent Work

Section 6: Workshop: Skills That Improve from Feedback

The agent creates and edits skills itself after complex tasks, and /learn builds a skill from a document or URL
The feedback loop: correct the agent's work and have it record the lesson in the skill for next time
Turn on human review before the agent changes a skill, and use the Curator to handle stale or duplicate skills
Install skills from the Skills Hub with care, check their content first and pin the important ones
Workshop: a reporting skill that measurably improves over 3 rounds of feedback, compared before and after

Section 7: Lab: Cron and Scheduled Jobs

Create jobs with /cron add, hermes cron create or a plain-language request
Schedule formats: every few hours, weekdays at 9am, cron expressions and one-off jobs
Deliver results to Telegram or a chosen topic, and script-only jobs that never call a model
Keep the gateway running, catch failed jobs and control the cost of frequent runs
Lab: a news or system status summary sent to the team chat every weekday morning

Section 8: Lab: Database MCP, CLI Tools, Plugins and Hooks

Connect a database MCP server with a read-only user and enable only the tools you need
Have the agent use CLI tools such as GitHub CLI through a skill to check issues and pull requests
Plugins bundle tools, skills, hooks and commands in one package, with commands to install and enable them
Hooks on agent events for an audit log of every tool call and alerts when something fails
Lab: ask for sales figures from a database in chat, with a hook that logs every command to a file

Section 9: Lab: Subagents, a Knowledge Base and Coding Agents

Delegate work to parallel subagents with delegate_task, each with its own separate context
Limit the number of subagents, isolate them in Git worktrees, and the idea of a Kanban board for multi-agent work
Use an Obsidian vault as a knowledge base the agent can search, cite and summarise
Hand coding tasks to a coding agent such as Claude Code or Codex, with Hermes tracking and reviewing the result
Lab: research a topic with 3 subagents at once and merge the findings into a new note in the vault

Section 10: Capstone: A 24/7 Agent That Runs Scheduled Work and Reports Through Chat

Choose a real brief or a sample one, such as a team assistant that summarises work, checks systems or watches a repository
Put it together: onboarding, a learning skill, cron, MCP or CLI tools and subagents where they fit
Check it against a safety, cost, logging and PDPA checklist before real use
Demonstrate it through Telegram, review it together and plan how to look after the agent after the course

Schedule & training options

For individuals — public rounds

No public rounds are open right now. Join the waiting list and we will contact you first when the next round opens, or ask us on LINE. Or call 02-570-8449 or 088-807-9770

Join the waiting list Ask about rounds

For organisations — in-house / private

Tailor the content to your team’s tools and projects
Your dates, at your office or live online
Quotation with tax ID for procurement

Corporate training quote

Instructors

Samit KoyomFounder & CEO, IT Genius Engineering

Frequently asked questions

How does this course build on Local AI Agent with Hermes and Ollama, and how does it differ from Personal AI Agents with OpenClaw and Codex and Local LLM Deployment with Ollama and vLLM?

This course builds on Local AI Agent with Hermes and Ollama, which covers installing Hermes locally, using Ollama and building a first skill. Here the basics are only briefly recapped before moving on to an always-on agent on a server, Telegram control, security, self-improving skills, cron, MCP, plugins, hooks and subagents. Personal AI Agents with OpenClaw and Codex uses OpenClaw and Codex, while Local LLM Deployment with Ollama and vLLM focuses on deploying and serving models, which can then act as an endpoint for the agent built in this course.

What do I need to prepare before the training day?

Bring your own laptop where you can install software, plus a Linux VPS you rent yourself, or use Docker or WSL2 on the laptop instead. You also need a Telegram account to create a bot and an API key from a model provider or a local model endpoint. You should be comfortable with a terminal, SSH and basic Docker, and have completed Local AI Agent with Hermes and Ollama or used Hermes Agent before. The institute does not provide machines, servers or accounts.

Are there costs beyond the course fee?

Hermes Agent is open source under the MIT licence and a Telegram bot is free to create, but VPS rental, pay-as-you-go cloud model usage and coding agent subscriptions such as Claude Code or Codex are paid by the learner. If you prefer not to rent a VPS you can use Docker or WSL2 on your own machine, and a local model can replace a cloud model in some labs. The institute does not provide any accounts or credits.

How safe is it to run an agent that can execute commands on a server all the time?

The risk is real and the setup must be tight. The course has a dedicated security section where you practise turning on command approval, setting a deny list, allow-listing who can message the bot, restricting writable folders, using Docker as a sandbox, keeping tokens secret, closing unused ports and logging every command. Personal data must follow PDPA and your organisation's policies. These settings greatly reduce risk, but they do not replace human oversight.

Can I use only local models, without any cloud model?

Yes, in many labs, by pointing Hermes at an Ollama or vLLM endpoint. However, an agent that works through many steps, calls subagents and edits its own skills needs a model that is strong at tool calling with a long context, and small models can be slower and make more mistakes, so results depend on the model and your hardware. In class you plan a main and a fallback model to suit the task and your budget.

Can Windows users take the course?

Yes. Windows users can install Hermes inside WSL2 or run it with Docker on their machine, or connect from Windows to a Linux VPS with SSH and VS Code Remote SSH. In class you set up WSL2 so the gateway starts with the machine, and learn how this differs from a server that is always on.