Subject: My attempt at an "OS-inspired" AI architecture Hi HN, I'm a Product Manager, not a systems engineer. I built AI Station Navigator as a proof-of-concept to solve a specific problem I faced: Context Pollution. When using AI agents for complex tasks, the context window gets cluttered quickly, causing the model to hallucinate or get confused. To solve this, I designed this project using a Computer Architecture Analogy. I treated the agent system like a traditional OS to better manage resources and isolation. Here is the mapping I used to design the system: -- CPU approx. LLM (Claude) The raw computing power driving the capabilities. -- Kernel approx. Orchestration Layer (Claude Code + CLAUDE.md) Handles intent recognition, task scheduling, and context management. -- Processes approx. Sub-Agents The core feature: Execution runs in isolated sub-agents. When a task is done, the sub-agent "dies," freeing up context. This prevents the main thread from getting polluted. -- Applications approx. Skills (GitHub Repos) "App Store-style" installation of tools via GitHub links. -- Drivers approx. MCP + Hooks Standardized interfaces for external tools and system automation. -- Runtime approx. Portable Environment Self-contained Python/Node environment (no installation hell). The Code: https://github.com/canishowtime/ai-station-navigator/ I'd love to hear your thoughts on this architectural approach.