125 lines
4.8 KiB
Markdown
125 lines
4.8 KiB
Markdown
---
|
|
title: I Built an Open-Source Agent OS — Here's the Architecture
|
|
published: false
|
|
description: "How I built Agentic OS: a locally-hosted multi-agent orchestration platform that coordinates opencode, Hermes Agent, and Gemini CLI into one dashboard with 16+ skills, cron scheduler, cost analytics, and persistent memory. Full architecture breakdown."
|
|
tags: [agents, devops, opensource, python, fastapi]
|
|
cover_image: https://raw.githubusercontent.com/modimihir07/agentic-os/main/docs/og-image.png
|
|
---
|
|
|
|
# I Built an Open-Source Agent OS — Here's the Architecture
|
|
|
|
Most AI agent tools work in isolation. You have a terminal for coding, a separate chat for research, another dashboard for monitoring, and yet another tool for memory. Each one is powerful alone, but together? Chaos.
|
|
|
|
That's why I built **Agentic OS** — a locally-hosted operating system for AI agents that coordinates three agents (opencode, Hermes Agent, and Gemini CLI) into one unified dashboard.
|
|
|
|
|
|
|
|
**GitHub**: [github.com/modimihir07/agentic-os](https://github.com/modimihir07/agentic-os)
|
|
|
|
## The Problem
|
|
|
|
I was using:
|
|
- **opencode** for code generation and DevOps tasks
|
|
- **Hermes Agent** for persistent memory and scheduling
|
|
- **Gemini CLI** for web research and analysis
|
|
|
|
Each was great at its job, but there was no central hub. No shared memory. No unified dashboard. No way to chain them together. Every session started from zero.
|
|
|
|
## The Solution: 7-Layer Architecture
|
|
|
|
Instead of a flat "three agents" approach, I designed a **7-layer architecture** that standardizes how agents interact:
|
|
|
|
```
|
|
Layer 7: Identity / Persona / Constitution
|
|
Layer 6: Self-Evolution + Capability Manager
|
|
Layer 5: Scheduler + Awareness + Health Guardian
|
|
Layer 4: Memory Graph + Memory Consolidation
|
|
Layer 3: Skills Hub + Eval + Learnings Loop
|
|
Layer 2: Business Brain + Context Folders
|
|
Layer 1: Agent Router + Standards + Profiles
|
|
```
|
|
|
|
### Layer 1: Agent Router
|
|
Routes tasks to the right agent automatically:
|
|
- **Code/DevOps task?** → opencode
|
|
- **Memory/Scheduling?** → Hermes
|
|
- **Research/Analysis?** → Gemini CLI
|
|
- **Complex multi-step?** → Chain: Gemini researches → opencode implements → Hermes monitors
|
|
|
|
### Layer 3: Skills Hub (16 Skills)
|
|
Each skill has a standardized structure:
|
|
```
|
|
skills/skill-name/
|
|
├── SKILL.md # Instructions
|
|
├── learnings.md # Accumulated knowledge
|
|
├── eval.json # Scoring criteria
|
|
├── score-history.json # Performance tracking
|
|
└── context/ # Task inputs
|
|
```
|
|
|
|
This means skills improve over time — the more you run them, the smarter they get.
|
|
|
|
### Layer 4: Persistent Memory
|
|
All agents read from a shared `brain/` folder at session start. Hermes provides SQLite FTS5 for full-text search across all past conversations.
|
|
|
|
## Feature Highlights
|
|
|
|
| Feature | What it does |
|
|
|---------|-------------|
|
|
| 3-Agent Engine | opencode + Hermes + Gemini CLI with auto-routing |
|
|
| 16 Skills | Executable packs with eval scoring |
|
|
| Cron Scheduler | APScheduler jobs (heartbeat, standup, audit) |
|
|
| Cost Analytics | Track spending across providers |
|
|
| One-Click Backup | Full tar.gz snapshots |
|
|
| Audit Trail | Every action logged |
|
|
| Prompt Library | 10 reusable templates |
|
|
| Dark/Light Theme | GitHub-style dark mode |
|
|
|
|
## Comparison with Julian Goldie's Claude Agent OS
|
|
|
|
I was inspired by Julian Goldie's "Agent OS: Claude + Hermes = Superpowers" video. Here's an honest comparison:
|
|
|
|
| Feature | Claude Agent OS | Agentic OS (Mine) |
|
|
|---------|----------------|-------------------|
|
|
| Core Agents | Claude + OpenClaw + Hermes | opencode + Hermes + Gemini |
|
|
| Cost | $20/mo (Claude) | **$0 (all free tiers)** |
|
|
| Stack | Next.js + Tailwind | FastAPI + vanilla JS |
|
|
| Memory | Obsidian (external) | **Built-in brain/ + SQLite** |
|
|
| Cost Tracking | None | **Built-in analytics** |
|
|
| Backup/Restore | None | **One-click** |
|
|
| Open Source | No | **MIT License** |
|
|
|
|
## Tech Stack
|
|
|
|
- **Backend**: FastAPI (Python)
|
|
- **Frontend**: Vanilla JS SPA (no framework)
|
|
- **Scheduler**: APScheduler
|
|
- **Memory**: SQLite FTS5 + Markdown files
|
|
- **Agents**: opencode, Hermes Agent, Gemini CLI
|
|
- **Cost**: $0 — all free tiers (Gemini Flash, OpenRouter free models)
|
|
|
|
## Quick Start
|
|
|
|
```bash
|
|
git clone https://github.com/modimihir07/agentic-os.git
|
|
cd agentic-os
|
|
chmod +x install.sh && ./install.sh
|
|
./start.sh
|
|
# Open http://127.0.0.1:8080
|
|
```
|
|
|
|
## What's Next
|
|
|
|
I'm planning to add:
|
|
- More skills (Kubernetes monitoring, CI/CD pipelines)
|
|
- Agent handoff protocol for complex multi-step tasks
|
|
- Web search integration across all agents
|
|
|
|
---
|
|
|
|
**Agentic OS** is open-source under MIT. Star it on GitHub if you find it useful!
|
|
|
|
👉 **[github.com/modimihir07/agentic-os](https://github.com/modimihir07/agentic-os)**
|
|
|
|
Questions? Comments? Open an issue or join the discussions on GitHub.
|