~6m21:19
Sharbel A.

100 Days With Hermes Agent in 21 Minutes

Jun 17, 2026

Read: ~6m · You save: 15 min

100 Days With Hermes Agent in 21 Minutes

Go beyond chatbot prompts. Learn how to build Hermes AI loops, Telegram workspaces, sub-agent teams, event triggers, and a mission control dashboard.

Many users see Hermes used to run businesses: finding leads, writing briefs, monitoring the internet, sending updates to a phone, and running background agents. A first installation often produces a much narrower result, such as summarizing a PDF. This gap can be closed by moving from one-off prompts to a structured system. The approach below is based on 100 days of working with Hermes. It covers repeatable loops, Telegram workspaces, sub-agent teams, event triggers, and a mission control dashboard.

Mistake 1: Using Hermes as a Smarter Search Box

Prompts such as "Give me ideas," "Summarize this," or "Research that" are useful. However, they do not compound. The answer is produced, and the work disappears after the conversation.

The better approach is to convert a task into a repeatable loop. The difference is clear in a YouTube example:

  • Basic prompt: "Give me YouTube ideas." Hermes returns a list of ideas.
  • Loop prompt: "Every time I ask for YouTube ideas, check my posted videos, my rejected ideas, my Notion content board, current demand, and competitor outliers. Then tell me the one video I should film next, why it matters now, what proof I need, and what the first 30 seconds should show."

The second prompt defines sources, rules, decisions, and an actionable output. The workflow is the unit of value, not the answer.

Identify manual decisions. Start by asking which decisions are made repeatedly by hand. These decisions usually mark the location of a loop. For example, one loop covers reaction opportunities. Without automation, the process involves scrolling X, scanning AI news, checking freshness, judging relevance, and deciding whether to post. Hermes can perform this process every day or every few hours. It can:

  • Check for fresh AI stories.
  • Query a source through an API provided by the user.
  • Check for product launches.
  • Filter for topics relevant to YouTube.
  • Prioritize items from the last hour.
  • Skip stale topics and duplicates.

Validate manually before automating. Each step of the loop should work manually first. For the YouTube ideation loop, the author confirmed that Hermes could:

  1. Fetch each component of the required data.
  2. Access the sources it needs.
  3. Use that data to generate video ideas.
  4. Produce ideas that are good and actionable.

Only after this validation was the loop scheduled.

Schedule with cron. A cron is a repeatable job that runs at a defined interval. Examples include:

  • Every morning: check what matters.
  • Every Friday: audit video ideas.
  • Every day: scan competitors.

The cron instructs Hermes to check specified sources, follow defined rules, ignore specified items, and send output on a schedule. The same pattern applies across roles:

  • A founder can summarize new leads each morning.
  • A creator can scan video opportunities each week.
  • A developer can open pull requests daily.
  • A salesperson can identify follow-ups that might otherwise be missed.

The value lies in converting repetitive manual work into a repeatable operating process.

Mistake 2: Keeping All Work in One Chat

An early setup placed YouTube ideas, X posts, business strategy, personal tasks, and general questions in a single chat. This made it impossible to determine what each conversation was for. The solution was Telegram topics. Each topic functions as a separate workspace.

The following topics were used:

  • YouTube: The only subject discussed in this topic. The sub-agent Nova also operates here.
  • X: Used to understand activity on X. It can answer questions such as whether the user was offline for the entire day, whether a refresher is needed, and whether Hermes can fetch sources for a tweet being written or published.
  • React: Sends reaction opportunities for topics to discuss on X. This is especially useful after a day offline or in meetings.
  • General: A catch-all topic for questions that do not fit other topics. The author strongly recommends this topic.

Each workspace has a defined job. When a message is sent in the YouTube topic, Hermes does not need to infer its mode, load a specific context, or store another type of context. It already operates within the relevant context.

This matters because the more capable an agent becomes, the more harmful mixed context becomes. A request for a video idea should not be processed alongside a grocery list, a code bug, or an unrelated email. Each request should be handled in the appropriate workspace, using that workspace's rules and agents.

Recommended starting setup: three topics.

  1. A main work topic.
  2. A content or research topic.
  3. An administrative topic.

Each topic should have one operating rule. Examples:

  • Content: Before suggesting anything, check what has already been posted, what has been rejected, what is currently trending, and whether the idea can actually be produced.
  • Administrative: Organize the task and draft the response, but request approval before sending anything externally.

Mistake 3: Relying on One Overpowered Agent

A single Hermes agent can perform many tasks, such as market research, competitor checks, Notion inspection, title generation, hook writing, and daily decision summaries. Assigning all tasks to one agent produces disorder quickly.

The better structure uses sub-agents. The main agent acts as the coordinator. Specialist sub-agents produce clean reports, and the main agent makes the final decision.

Example: selecting the next video.

  • Sub-agent 1 checks the posted and filmed pipeline.
  • Sub-agent 2 checks competitor outliers.
  • Sub-agent 3 checks current demand.
  • The main agent reads all three reports and decides which video to film.

This structure produces cleaner work, cleaner context, and more reliable decisions. Once the workflow is defined, it can be turned into a repeatable cron or a repeatable skill. The agent then deploys the defined sub-agents without starting from scratch each time.

Specialization over quantity. More agents do not automatically produce better results. More focused agents do. Each agent requires a specific lane, with its own tools, memory, and permissions. A YouTube agent should not match a coding agent, and a coding agent should not match a personal admin agent.

Two specialist agents are referenced:

  • Nova: The YouTube agent. Nova is trained to find and check YouTube trends, compare video outliers and competitor posts, identify the best video ideas, and support research and ideation.
  • Sage: The X and content strategy agent. Sage works on the same principle as Nova, focused on X.

Both agents are open source and free to download and install. Their links are provided in the resources section below. Users can copy their structure or build their own specialized agents.

Mistake 4: Relying Only on Time-Based Triggers

Cron handles time. Webhooks handle events.

  • A cron repeats a task after a defined interval.
  • A webhook runs a task when a specific event occurs.

Webhooks make Hermes responsive to real activity.

Example: a Notion content board. When a video idea is moved to the "Film" tab, the change triggers a webhook. Hermes then:

  • Checks whether a similar video has already been posted.
  • Assesses current demand for the idea.
  • Identifies the strongest title.
  • Recommends whether to film the video now, film it later, or abandon it.

In this setup, Notion functions as a control surface, not only as a database. A status change becomes the trigger for a chain of work.

Other trigger examples:

  • A form is submitted: Hermes researches the lead.
  • A GitHub pull request opens: Hermes reviews the risk.
  • A customer pays: Hermes adds the payment to the daily summary.
  • A file is uploaded: Hermes extracts the relevant content.
  • An email is received: Hermes reads and summarizes it and drafts a possible reply.

Setup. A webhook route is created in Hermes. The user can ask Hermes how to create webhook triggers. With a capable model connected to the application, Hermes can perform the configuration. The user specifies the required outcome, and Hermes runs the workflow each time the event occurs.

Polling cron as a fallback. If an application does not support a clean webhook, a polling cron provides a close alternative. The cron checks for a change at a fixed interval, for example every five minutes. It compares the current state with the previous state and acts only when the defined event has occurred.

Principle. The user should not remain the trigger that waits to request each action. Tools and data sources, such as email and databases, should trigger Hermes. The system then operates without continuous user prompting.

Mistake 5: Managing a System Only Through Chat

Chat works well for commands. It does not provide sufficient visibility into a running system. Active work requires clear information about what is running, what has finished, what has failed, and which outputs need attention.

Mission control is the dashboard layer of the system. The author's implementation is a live Vercel website connected to the user's workflows. Its features include:

  • Authentication: Access is protected with Google authentication through Gmail. This prevents unauthorized users from locating the site and controlling the system.
  • Mobile access: The dashboard can be opened on a phone. Hermes can be controlled from a phone, from any location, without requiring access to the local machine. This applies even when Hermes runs on a local instance.
  • Visibility and approval: Users can see what Hermes is doing, which allows them to delegate more work with confidence.

A toy setup hides all activity in chat. A working setup provides command, visibility, and approval.

Recommended Build Order

For a new installation, the recommended sequence is:

  1. One repeatable loop.
  2. One Telegram workspace.
  3. One specialist agent.
  4. One scheduled cron.
  5. One event trigger.
  6. Mission control, once the system becomes too large for chat alone.

Each layer should be added only after the previous one works reliably. The goal is not to use every feature. The goal is to convert the user's workflow into the system.

Summary

FromTo
AnswersLoops
One chatSeparate workspaces
One agentSpecialist sub-agents and a team
Manual requestsCron schedules and event triggers
Hidden activityMission control dashboard

Key principle: Teach Hermes how the work runs. Instructing the agent about the process, rather than only asking it for help, is the source of leverage.

Resources

  • Nova (YouTube agent) and Sage (X and content strategy agent): open source, available on the author's GitHub.
  • Mission control template: open source and free to download. A link is provided in the resources section.

Mistake One: Using Hermes Like a Search Box

Using Hermes only for one-off questions wastes its potential; the work disappears after each answer. The upgrade is turning tasks into repeatable loops.

  • The first mistake is using Hermes like a smarter search box for one-off prompts such as 'give me ideas', 'summarize this', or 'write a draft'.
  • One-off answers do not compound because the work disappears after the answer.
  • Instead of asking what Hermes can answer, ask what decisions you keep making manually.
  • Example loop: a YouTube idea request that checks posted videos, rejected ideas, the Notion content board, current demand, and competitor outliers, then names the one video to film next, why it matters now, what proof is needed, and what the first 30 seconds should show.
  • Example loop for AI news: check fresh AI stories every couple of hours, use an API for sources, filter for YouTube-relevant topics, prioritize the last hour, and skip stale or duplicate topics.
  • Pro tip: make the loop work manually first, confirming each data source, the data usage, and the quality of the output, before scheduling it.

Scheduled Loops with Cron

A cron is a scheduled repeatable job. Once a manual workflow is proven, scheduling it turns Hermes from answering requests into reporting back.

  • A cron is a repeatable job that runs on a schedule, such as every morning, every Friday, or every couple of hours.
  • Examples: every morning check what matters, every Friday audit video ideas, every day scan competitors.
  • Examples by role: a founder summarizes new leads daily, a creator scans video opportunities weekly, a developer opens PRs daily, a salesperson finds forgotten follow-ups.
  • The value is turning daily grunt work into a repeatable operating system.

Chats Become Workspaces

Keeping all work in one chat creates noise. Telegram topics act as separate workspaces, each with its own job and rules, so Hermes does not have to guess its mode.

  • The first setup was one giant chat mixing YouTube ideas, X posts, business strategy, and personal tasks, which became impossible to organize.
  • The fix was Telegram topics, with each topic treated as a separate workspace.
  • Suggested topics: YouTube (content ideation, where sub-agent Nova lives), X (understanding what is happening on X, refreshers, sourcing for tweets), General (catch-all), and React (sends reaction opportunities for things to talk about on X).
  • Each Telegram topic is a separate workspace with its own job.
  • Messaging Hermes in the YouTube topic means it is already in the right context and does not need to load or store another context.
  • Recommended start: three topics, one for main work, one for content or research, and one for administrative tasks.
  • Example operating rules: for content, check posted content, rejected ideas, trending topics, and feasibility before suggesting anything; for admin, organize tasks and draft responses but ask before sending anything externally.

Mistake Three: One Agent Doing Everything / Sub-Agents

One overpowered agent gets messy. A main agent coordinating specialist sub-agents produces cleaner context and better decisions.

  • Giving one Hermes agent everything (research, competitors, Notion, titles, hooks, daily decisions) becomes messy quickly.
  • The main agent acts as operator, coordinating specialists and making the final call using their clean reports.
  • Example for picking the next video: one sub-agent checks posted and filmed pipeline, one checks competitor outliers, one checks current demand; the main agent reads the reports and decides what to film.
  • Once a workflow is defined, it can be turned into a repeatable cron or skill so Hermes does not start from scratch each time.
  • More agents does not automatically mean better; more focused agents with specific lanes means better.
  • Different jobs need different tools, memories, and permissions.
  • Nova is the speaker's YouTube agent, trained on finding and checking YouTube trends, comparing video outliers and competitor posts, finding video ideas, and assisting with research and ideation. It is open-source and free; the link is in the description.
  • Sage is the speaker's X and content strategy agent, similar to Nova but dedicated to X. It is also open-source and free; the link is in the description.

Event Triggers: Crons and Webhooks

Cron handles time-based triggers; webhooks handle event-based triggers. Event triggers make Hermes proactive, reacting when something happens in connected tools.

  • Cron handles time: repeatable jobs at set intervals.
  • Webhooks handle events: they run when something happens.
  • Example: when a video idea is moved to the 'File' tab in the Notion content board, a webhook triggers Hermes to check for similar past posts, current demand, the strongest title, and whether to film now, later, or abandon it.
  • Notion becomes the button that starts the workflow, not just a database.
  • Other trigger examples: a form submission triggers lead research; a GitHub PR opening triggers a risk review; a customer payment adds to the daily summary; a file upload triggers extraction of useful parts; an email triggers a summary and a draft reply.
  • Setup: create a webhook route in Hermes; you can ask Hermes how to create webhook triggers and it can guide or perform the setup, given a capable model connected to the app.
  • If an app does not support clean webhooks, use a polling cron that checks every few minutes, compares changes, and acts only when the right event occurs.
  • The lesson: stop making yourself the trigger; let tools, emails, and data trigger Hermes.

Mission Control Dashboard

Chat is good for commands but poor for visibility. A mission control dashboard shows what Hermes is running, finished, or failed, and allows control from anywhere.

  • Chat is good for commands but bad for visibility.
  • Mission control is the dashboard and information layer, with visibility into what is running, what finished, what failed, and what is needed.
  • The speaker's mission control is a live Vercel website connected to his workflows.
  • Access is protected by Google/Gmail authentication so random people cannot log in.
  • It can be opened on a phone, allowing control of Hermes from anywhere, including while traveling, at the gym, or on holiday.
  • Approval and visibility give confidence to let Hermes do more.
  • The speaker's mission control template is open source and free, with a link included below the video.

Getting Started: Build Order

The recommended build order is incremental: start with one loop and add one layer at a time, from workspaces to sub-agents, crons, webhooks, and finally mission control.

  • Build order: one boring loop, one Telegram workspace, one specialist agent, one scheduled cron, one event trigger, then mission control once the system is too big for chat.
  • Answers become loops, chats become workspaces, one agent becomes many agents, manual work becomes triggers, and invisible work becomes mission control.
  • Specialist agents Sage and Nova and the mission control template are open source on the speaker's GitHub, with links below the video.
  • Teach Hermes how your work runs rather than just asking it for help.