I walked into this with exactly zero Rust experience and a stubborn belief: building an agent shouldn’t require a megachurch of frameworks. Twenty-four hours later, I shipped a Rust-native AI Agent desktop for non-dev that compiled to a 16MB tiny binary, a simple while loop with Skills, MCPs supported. And the most important part: BYOK (Bring Your Own Key) and even local models!

Call it technical minimalism.

The Stack

• Rust (from zero to shipping in one day)
• Long term memory: File system as persistent, queryable context
• Loop: A single while loop coordinating I/O and tool calls
• Tools: Built-in file tools (read, write, bash) + MCP-compliant providers (extensible)
• Sandboxing: A simple docker container would isolation most risks.

Secret Sauce

• Keep state in plain files and directories.
• Keep orchestration in a simple loop, not a framework
• Teach the AI to read and write its own operating context, in files.

What I learnt: 1) You don’t need an agent framework Agent “framework” is almost always over abstraction and adds complexity.

• Just a simple while loop:

1. Observe: read inputs, logs, and task files
2. Decide: ask the model for the next atomic action
3. Act: call a tool (MCP) or write to disk
4. Reflect: append outcomes to a local journal

2) The File System is all you need Vector stores and RAG farms are great—until they become speculative overhead.

• Everything as Files: Tasks, plans, diffs, decisions, and logs live as plain text and JSON.
• Modern RL training methods have trained models to work super well with Linux files systems.
• Using the FS as the “long term momery” makes persistence trivial and recoverability obvious.

3) The Surreal Mirror: AI Coding reproduces itself. Of course, I use an AI agent (Claude Code) to work for me. The Open Cowork is basically a GUI with a CC-like agent behind it. It’s super interesting to see CC can reproduce itself in Rust.

4) Security as a First-Class Citizen Agents are powerful, but power without boundaries is a liability. They are potentially vulnerable to bad actors and prompt injection risks or even bugs. For non-devs, security isn’t just a feature—it’s the foundation. By implementing hard sandboxing (like Docker) and strictly whitelisting system commands or network calls, we ensure the agent remains a helpful tool, not a system threat.

The entire codebase can be summarized in the loop below:

Of course it’s still early, any feedback is welcome!