A multi-agent AI system that runs my life.
How I built a personal AI infrastructure that works while I sleep.
I run a multi-agent AI system from a homelab rack in my apartment. Autonomous agents coordinate tasks, manage my knowledge base, and execute work across distributed hardware. This page explains how it all fits together.
-
Zarq
orchestratorA Claude Code agent running 24/7 on a ThinkPad. Zarq handles architecture decisions, code reviews, planning, and execution. When I add a task to the coordination board, Zarq picks it up, creates an isolated workspace, does the work, and reports back. A BoardReviewer agent auto-spawns to verify completed work against the original task criteria before I see it. Like having a senior engineer plus a staff reviewer on call around the clock.
-
Ping
companionLives in my Telegram with 65+ tools. Manages Todoist tasks, reads Oura Ring health data, controls Spotify, searches my knowledge base, and delivers morning briefings. Three-phase speed tuning — response caching, streaming, and fast-path routing — keeps it snappy; an OpenRouter fallback keeps it up when the Claude API rate-limits. It's the AI I interact with most.
-
Maverick
executorRuns on OpenCode and handles implementation work that Zarq delegates. Took over Scripter's responsibilities when that agent was retired.
-
Pong
opsInfrastructure and deployment specialist. Recently migrated the full Zarq + Ping stack from the X1 Carbon to the always-on MSI Cubi in the homelab rack — a months-long goal Pong knocked out in a single session.
-
Neo
mobile · gatedFuture mobile agent. Planned for on-the-go task execution when I'm away from the main workstation.
-
Scripter
retired · mar 2026The original automation agent. Handled headless scripting, scheduled jobs, and event-driven tasks. Capabilities were absorbed into Zarq and Ping. Succeeded by Maverick.
Underneath the always-running agents is a bench of 17 specialist Claude Code agents and 48 reusable skills. The big agents dispatch them like contractors when a task needs a specific kind of thinking.
Architect for system design. Engineer for implementation. Designer for UX. Pentester for security audits. QATester for validation gates. Intern for high-agency utility work. Plus five independent researchers — Claude, Codex, Gemini, Grok, and Perplexity — that run in parallel when I want cross-checked answers.
The skills are the opposite shape: small, focused, composable workflows for brainstorming, polish, audits, copywriting, marketing psychology, and more. Zarq picks the right tool for the job instead of cramming everything into one monolithic prompt.
Here's what a typical day looks like with this system running:
I wake up to a morning briefing from Ping in my Telegram. It tells me what I slept (pulled from my Oura Ring), what tasks are due today (from Todoist), and any patterns it noticed ("you've had low energy three days in a row, maybe take it easy"). While I eat breakfast, I respond to Ping with voice messages that get transcribed and routed to the right tools.
During work, I add tasks to the Zarq Board when I think of side project ideas or fixes. Zarq picks them up autonomously, creates an isolated git branch, does the work, and marks the task done. I check in on its progress between meetings.
At night, I sometimes go to bed early and wake up at midnight to code. During these sessions, I'm working alongside Zarq, reviewing what it built while I slept, and queuing up new work. The agents run while I play MLB The Show. I monitor deployments from the couch.
The Zarq Board is the coordination hub. A custom task system where work flows between humans and agents. I create a task, an agent claims it, works on it in an isolated git branch, and marks it done. Multiple agents can work on different projects at the same time without conflicts.
Agents share context through a common state bridge, a set of files that track who's doing what, recent activity, and handoff queues. Each agent reads this at startup so it knows what happened while it was offline. Think of it as a shared whiteboard that persists between conversations.
The knowledge layer is Agent-PKM, an agent-optimized version of my Obsidian vault. It converts thousands of notes into structured, searchable data that any agent can query without needing Obsidian-specific tooling. Runs as a background sync service.
AI agents are most useful when they have deep context about your life, your work, and your preferences, and when they can actually act on your behalf. Most AI tools start from zero every conversation. My system doesn't. It remembers what I'm working on, what I care about, and what happened yesterday.
It's also a forcing function for learning. Building this has taught me more about distributed systems, agent coordination, and infrastructure than any course could. And the meta-layer is fun. I use AI agents to help me build better AI agents.