The speaker, a creator and developer, shares hard-won lessons from 100 days of using Hermes Agent, transforming it from a simple query tool into a personal operating system. The video bridges the gap between casual AI use and building a systematic, autonomous workflow. The core thesis is that true leverage comes not from asking questions, but from teaching the agent how your work runs.
The journey from random questions to a self-running system is broken down into five distinct phases. Each phase represents a shift in mindset and capability, moving from one-off tasks to a durable, scalable infrastructure. The speaker emphasizes that every powerful system starts with a single, boring, repeatable task.
Phase 1: Repeatable Loops (Mistake #1: Using Hermes like a search box) The first critical error is treating Hermes as a smarter search engine for one-off answers, which do not compound. The upgrade is to identify decisions you make repeatedly and turn them into a repeatable workflow.
- 🧠 Shift from answers to actions: Instead of asking for "YouTube ideas," create a prompt that checks posted videos, rejected ideas, a Notion board, current demand, and competitor outliers before outputting a single actionable video idea with a reason and first 30 seconds.
- 🔁 The manual-first rule: Verify that Hermes can fetch every individual piece of data from its required sources before combining them into a workflow. Once confirmed, schedule it as a cron (a repeatable job).
- 📈 Compounding value: The value is not a single task, but turning grunt work into a repeatable operating system for founders (summarizing leads), creators (scanning video opportunities), or developers (opening daily PRs).
Phase 2: Telegram Workspaces (Mistake #2: One giant junk drawer) When all work lives in one chat, everything becomes noisy and context is lost. The fix is not a more complicated app, but using Telegram topics to create separate, dedicated workspaces.
- 🗂️ One topic, one job: Create separate topics for YouTube, X (Twitter), general questions, and reaction opportunities. Each topic has a single purpose, so the agent does not need to guess its mode or load irrelevant context from a grocery list or code bug.
- ⚙️ Workspace rules: Give each topic a simple operating rule. For a content topic, the rule might be: check what I already posted, what's trending, and whether I can make it. For admin: organize, draft, but ask before sending.
- 🧱 Practical setup: Start with three topics: main work, content/research, and administrative. This immediately improves output quality by eliminating messy context.
Phase 3: Sub-Agents (Mistake #3: One agent trying to do everything) A single agent doing everything becomes messy. The better pattern is a main agent that coordinates with specialized sub-agents, much like a manager with a team.
- 👥 Specialists over generalists: For a video pick, one sub-agent checks the posted pipeline, another checks competitor outliers, and a third checks current demand. The main agent reads clean reports and makes the final call.
- 🎯 Specific lanes: More agents do not mean better; more focused agents do. Each needs a specific memory, permission, and toolset. The speaker open-sourced two examples: Nova (YouTube agent) and Sage (X and content strategy agent).
- 🧩 From ad-hoc to skill: Once a workflow is defined, it can be turned into a repeatable cron or skill. When asked for a YouTube idea, the system deploys the three sub-agents automatically without starting from scratch.
Phase 4: Event Triggers (Crons & Webhooks) The system needs to know when to act. Crons handle time-based repetition, but webhooks make the system feel alive by reacting to events.
- ⏰ Crons for time: "Do this every morning" or "scan competitors every Friday."
- ⚡ Webhooks for events: "Do this when something happens." For example, when a video idea is moved to the "film" tab in Notion, a webhook triggers Hermes to validate it, check for similar content, and find the strongest title.
- 🔗 Real-world triggers: A form submission triggers lead research; a GitHub PR opens triggers a risk review; a file upload triggers data extraction. If an app lacks webhooks, use a polling cron that checks every few minutes and acts only when the right change is detected.
Phase 5: Mission Control (Mistake #5: Managing a system from a chatbot) Chat is good for commands but terrible for visibility. A real system needs a dashboard to see what is running, what finished, and what failed.
- 🖥️ The dashboard layer: The speaker built a live Vercel website as "Mission Control," authenticated via Google, so it can be opened from anywhere in the world on a phone or laptop.
- 🎛️ Command and visibility: A mission control provides the confidence to let the agent do more because you can actually see what it is doing. It is the difference between a toy setup (hiding everything in chat) and a real setup (command, visibility, and approval).
Final Takeaway: The Path from Chatbot to Operating System
The point is not to use every feature, but to transfer your workflow into the system. Start by picking one task you repeat over and over and turn it into a loop. From there, add one layer at a time: workspaces, sub-agents, crons, webhooks, and finally mission control. Do not just ask Hermes to help; teach it how your work runs. That is the leverage. Answers become loops, chats become workspaces, one agent becomes a team, manual work becomes triggers, and invisible work becomes your mission control.