AgentOS
Self-hosted AI agent operating system
A robust platform for orchestrating autonomous workflows, tools, workers, and persistent memory using a Planner/Supervisor architecture.
Problem
I needed a way to orchestrate complex, multi-step AI tasks that transcend a single context window. Existing frameworks were either too rigid or designed as opaque SaaS products.
Approach
I built AgentOS to fully own the execution layer, memory, and observability. It operates on a Planner/Supervisor architecture, distributing tasks across dedicated worker queues while maintaining state through a unified memory layer.
What shipped
- Decoupled planning phase from execution significantly reduces hallucination cascading.
- Combined vector search with structured SQL metadata yielding much higher retrieval precision.
- Built a reliable observability pipeline for non-deterministic AI outputs.
Stack
Next.js
Python
FastAPI
Postgres
Redis
Docker