A developer with no Rust experience built a native AI agent desktop in 24 hours, producing a 16MB binary. The project uses a single while loop for orchestration, treats the file system as long-term memory, and supports BYOK and local models. It implements Docker sandboxing for security, aiming for technical minimalism over complex frameworks.

This approach rejects sprawling agent frameworks, arguing they add unnecessary abstraction. The core loop follows observe-decide-act-reflect, reading inputs and writing outputs directly to plain files and JSON. By using the OS file system for persistence instead of vector databases, it reduces overhead and improves recoverability, aligning with how modern models are trained.

The project emerged from using AI coding assistants like Claude Code, which effectively helped reproduce its own functionality in Rust. For non-developers, security is foundational; hard sandboxing and command whitelisting prevent the agent from becoming a system threat. The codebase remains simple, focused on a single coordinating loop rather than a monolithic framework.