Picture this. You ask a bot in your Discord to build a new feature for your app, then walk away. A few minutes later, it pings you back with a working build. That is exactly what I watched happen, and I could not look away.
The person who pulled this off is Tina Huang, an ex-Meta data scientist who now runs a company called Lonely Octopus. She built a full multi-agent system she calls HermesOS, and it does way more than write code. It acts as her software engineer, her health coach, and even the COO of her startup. I think this is one of the clearest real-world looks at what a personal agent army can actually do.
Let me break down what she showed off.
🛠️ What’s new: three agents doing real jobs
HermesOS is not one bot. It is a crew of specialized agents, each with its own job and its own chat home. Here are the three the creator uses most:
- Coder. She talks to it in Discord. She asked it to add skin-switching to her little pixel Pomodoro timer app, and it drafted a full spec, waited for her approval, then built it. Minutes later the avocado timer could swap into a cat, a dog, and a slightly creepy tomato.
- Lifebot. This one lives in Telegram and acts as her productivity and health coach. It pulls from her Pomodoro logs, her custom to-do app, and her Apple Health step data to give real advice.
- Taco Bot. This is the company brain. It sets reminders, scans Discord and GitHub to report what the team is working on, and drafts operations guides by digging through Notion, Slack, and Google Drive.
🔀 The twist: the bot you talk to is not the bot that works
Here is the part that surprised me. The Hermes bot she chats with in Discord runs on a local model on her Mac Studio, a Qwen 3 model with billions of downloads. But that local model does not write the code itself. It pokes Claude Code in the background, hands off the whole spec, and lets Claude do the heavy building. When Claude finishes, it tells the local model, which then tells her it is done.
Why the roundabout setup? Because she pays for a Claude Max subscription and wants to actually use those tokens, but Anthropic does not let you plug that subscription in directly. So the original poster built a clever workaround. It sounds complicated, and it kind of is, but it means her cheap local model handles the chatting while her premium subscription handles the real work. Smart use of what she already pays for.
📋 The mini-workflow: how to think about building your own
If you want to copy the idea, here is the pattern I pulled from her setup:
- Pick one job. Do not build a do-everything bot. She has one agent for code, one for health, one for ops.
- Give it a home. Each agent gets a chat channel she already checks, like Discord or Telegram. No new app to babysit.
- Feed it real data. Lifebot is only smart because it can read her Pomodoro logs, her to-do lists, and her step count from an Obsidian vault. The advice is only as good as the data behind it.
- Let a strong model do the hard part. A small local model can route and chat, while a bigger model builds or reasons.
- Add quiet cron jobs. She has an alerts channel where agents check in daily or weekly to confirm nothing is broken and no security issues popped up. Not flashy, but it keeps the whole thing alive.
What blew me away was the Lifebot coaching. It told her that her 9 to 11 a.m. slot was a dead zone, and that she did not need more breaks, she needed one more focus session per day. That kind of pattern-spotting on your own data is hard to get from a generic app.
💡 Pro tips from her build
- Match your break length to your habits. She learned 5-minute breaks turn into procrastination and 15-minute breaks mean she never comes back, so Lifebot now tells her to take exactly 10.
- Swap models freely. She runs DeepSeek, Gemini Flash, Qwen, GLM, and GPT depending on the task, and slowly moves company data to local models for privacy.
- No dedicated machine? She runs a second team instance on a VPS through Hostinger, which even has a pre-made Hermes template so you skip the manual config. One rented box, as many agents as it can handle, no per-agent fees.
- Use it for onboarding. New team members get access to Taco Bot so they can just ask how the company does things instead of reading a wiki.
The COO angle is the sleeper hit here. She said the COO role was the hardest seat to fill for years, and Taco Bot now drafts the ops guides, documents the release checklists, and answers team questions from all their tools. For a 10-person team, that is a huge lift off her plate.
🚀 Want the full tour?
This is one of those videos where seeing it move makes it click. She walks through each agent live, shows the actual build finishing, and puts spec summaries on screen for every piece. If you have been curious about turning AI into a real teammate instead of a chat window, go watch the full breakdown and steal a few ideas for your own setup. 🐙