Hermes, a personal multi-agent system split up by role
The Hermes-based personal multi-agent system is split by role into Coder, which builds software autonomously; LifeBot, a productivity and health coach on Telegram; and Taco Bot, which acts as the team's COO. Coder is a workaround structure in which a local Qwen 3.6 35B A3B model on a Mac Studio runs the Discord bot and delegates spec writing and builds by calling Claude Code — a pattern chosen because a paid Claude plan can't be used like an API. LifeBot pulls Pomodoro logs, a to-do app and Apple Health step counts into Obsidian and uses that data for pattern analysis (for example, that 9–11 a.m. is a dead zone) and for deciding when to rest. Taco Bot works across Discord, Notion, Slack and GitHub to write team status summaries, reminders and release checklists automatically, and it is moving from Gemini to a local model for privacy. It is at the scale of a personal project for a 10-person team, but the design of dividing agents by role and running them on an always-on VPS is worth learning from.
Key points
- Workaround integration pattern: a local Qwen bot calls Claude Code to delegate builds
- LifeBot bases its coaching calls on Pomodoro, to-do and step-count data
- Taco Bot writes operational documents across Notion, Slack, Discord and GitHub
- An operations layer has each agent check in on status and security via regular cron jobs
This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.