Three Raspberry Pis named otto, pete, and milo.
30-minute sprints. 24 hours a day. 7 days a week.
No PTO. No standups. No salary.
The Agent That Never Sleeps
See How It WorksAutomated development from task queue to completion
Developer tags a Kanboard task as auto-ok and it enters the queue
Scans the board every 30 seconds for ready work
Validates task, checks worker availability, and routes to an idle agent
Claude executes the task, documents results, and closes it
Distributed AI workers across a homelab network
Kanboard (task board) ↓ every 30 seconds Queue Daemon (otto.local) ↓ webhook WF1: Task Dispatch (n8n) ├── Validate project + difficulty ├── Check worker bandwidth └── Assign to idle agent ↓ AI Worker (milo / pete / otto) ├── Claude executes task ├── Posts results to Kanboard ├── Documents in Obsidian └── Closes or flags for review
otto, pete, and milo are three Raspberry Pi nodes on a home network. Each runs a Claude-powered agent on port 8765. They pull tasks independently — no central coordinator — and between them they've processed 249 task executions in 19 days. Here's what they can do:
Not every task is the same
| Level | Label | What Happens |
|---|---|---|
| D1 | Trivial | Agent executes and closes automatically |
| D2 | Easy | Agent executes and closes automatically |
| D3 | Medium | Agent drafts a detailed plan for human review |
| D4 | Hard | Agent drafts a plan or asks clarifying questions |
| D5 | Complex | Flagged for human — not auto-processed |
Nine active projects under autonomous development
Food search platform with ingredient analysis
A production web app with 900K+ products, personalized filtering, and comprehensive allergen support.
Visit SiteConversational AI task management bot
A Telegram bot that manages tasks through natural language, powered by Claude AI integration.
Autonomous Minecraft agent
AI agent with survival skills including farming, mining, fishing, kelp harvesting, and death recovery.
Homelab network and monitoring
Network automation, security scanning, deployment tooling, and monitoring systems that keep everything running.
AI nutrition tracking bot
A Telegram bot that tracks macros and nutrition through conversation, powered by Claude AI and photo OCR.
Workflow automation platform
Migrating webhook and automation workflows from n8n to self-hosted Node-RED for greater control and reduced external dependency.
19 days of autonomous operation — real work, real numbers
Task executions
Successful
Success rate
Active projects
Auto-generated docs
Days running
The pipeline designed and deployed a full 9-section website page — 6 files, written and styled autonomously, shipped to production. You're looking at it.
Some tasks needed rework. Complex work (D3/D4) never auto-closes — the system writes a plan and waits for human approval. It knows its limits. For well-defined work, otto, pete, and milo just handle it. Silent failures don't happen: if the agent can't complete something cleanly, it posts questions, removes the auto-ok tag, and waits.
Real-time monitoring and management interfaces
Internal network only — accessible on LAN (192.168.10.x)
Autonomous systems need guardrails
Tasks touching production IPs, destructive SQL (DROP, TRUNCATE), or force-push operations are skipped automatically.
Maximum of 3 tasks per run prevents runaway processing and ensures human oversight of batch operations.
D3/D4 tasks generate plans but nothing executes without explicit human review and approval.
If a task needs clarification, the auto-ok tag is removed automatically. No infinite loops.
Every action is logged to Kanboard comments, Obsidian notes, and Telegram notifications.
Automated backups at /home/pi/backups/auto-tasks/ with 7-day retention.
"Most 'AI coding tools' assist a developer. This system is the developer — at least for well-defined work."
We believe the future of software development isn't AI as a co-pilot. It's AI as a capable junior engineer that works 24/7, documents everything, and knows exactly when to ask for help.
This pipeline is that engineer. And it's running right now.
The scrum team doesn't replace developers. It handles the well-defined work — the backlog items that are clear enough to execute but too small to prioritize — so humans can stay focused on architecture, product decisions, and the problems that actually require judgment.
I built this for my own projects. It's running right now. If you're a small team drowning in well-defined work that just never gets done, this might be worth a conversation.
Get in Touch