No security gates: 14-year-old Hees creates tool for AI agents to band together like a team

A team that acts on its own

Autark is designed to coordinate AI agents that can take action when a schedule, timer, task, message or event activates them. The tool can also let agents plan what to do next, rather than waiting for each instruction.

Model choices and limitations

Autark can connect through command-line tools for OpenAI Codex, Claude Code, Gemini CLI, GitHub Copilot CLI and Kimi Code. It also supports local models. The project remains under active development. Its website says support for models beyond Codex and Claude Code is built in, but has not yet been tested in a live run.

Users must set boundaries

Autark does not provide sandboxing, approval prompts or permission gates. Its agents run with the privileges of the machine’s user, and stated limits may not prevent them from attempting conflicting actions.

The tool does not stop agents from deleting files, sending unsolicited emails or spending available money. Hees acknowledges the risks, and the article advises using Autark on a machine without anything users would not want to lose.

Protection for stored secrets

Autark’s website says passwords and keys are encrypted and referenced by IDs, keeping their values out of prompts. It also says its servers do not store user data.

A meeting server relays encrypted traffic between devices. These protections do not remove the need for users to establish their own safeguards around what agents can access and do.