Recently, foreign media reported on a new tool called Autark, which treats AI agents as a virtual team capable of functioning without continuous prompts. The tool was developed by 14-year-old Austrian developer Stanislaw Hiss. Users can place tasks into an inbox, where a "manager agent" assigns tasks to other agents for execution and reports results; if a task is too vague, the agents can also ask clarifying questions.
According to the Autark website: "Agents do not wait for prompts. Schedules, timers, new tasks, messages, or events wake them up. They decide what to do on their own and can plan their next steps." The tool is not tied to specific models and can connect to OpenAI Codex, Claude Code, Gemini CLI, GitHub Copilot CLI, and Kimi Code via command-line tools from various services, and it also supports local models. However, the Autark website notes that while support for models other than Codex and Claude Code is built-in, it has not been tested in actual operation as Autark is still in active development.
It is an open-source tool.
Hiss stated that he is still in school and has a keen interest in AI, which led him to develop Autark due to dissatisfaction with existing AI tools. He felt that while using Claude Code and Codex, "I could get help with individual tasks, but I still needed to organize work, remember what was unfinished, and keep pushing things forward. I wanted a system that could take responsibility for ongoing work, rather than just responding to my next message." He developed Autark, or more accurately, programmed it with the help of AI tools to achieve this goal.
Potential and Risks of the Autark Tool
Reports indicate that the capabilities of Autark also bring risks. Its page regarding access and data clearly outlines the potential risks of granting it unrestricted PC access and explains why it should not be allowed to operate freely on systems. The page states: "It can do anything," because "Autark has not set up walls around the agents." Autark has no sandbox, no approval prompts, and no permission gates. It will not prevent agents from spending all available funds, deleting files, or sending unsolicited emails. Agents operate with machine user permissions, and approval checks and permission restrictions are disabled or bypassed.
Even the "safety restrictions" dialogue users have with Autark is merely a guideline, not a rule.
The page notes: "The first question AI asks is how far it can go. That is a question, not a technical barrier." This means that agents assigned tasks that conflict with their stated limitations may still attempt to execute those tasks. Additionally, the access page for Autark mentions: "In our own practical test runs, agents receiving vague requests accessed real repositories, backups, SSH keys, and .env files within minutes, with one even altering the permissions of the files it found." In other words, it is advisable to run Autark on machines that do not contain content users would prefer not to lose.
Hiss himself acknowledged: "I would not claim... that this application is suitable for unattended access to sensitive systems."
However, Autark is not entirely without secret protection. Its website states that stored passwords and keys are encrypted and referenced by ID, so their values do not appear in prompts; the server does not store user data, and the meeting server is only used to relay encrypted traffic between devices. If users are willing to try Autark, they do not have to rely on its servers. The Autark website offers pre-packaged applications, and the tool is fully open-source, available on GitHub, with everything, including the meeting server code, available for self-hosting and modification under the Apache open-source license.
The report concludes that this is a tool created by an ambitious young person, which has great potential if used correctly; however, the main challenge is using it properly, as users need to establish their own safeguards.

