A 14-year-old Austrian developer has created Autark, an application that organises AI agents into a team capable of continuing work between conversations, offering a glimpse of more autonomous software while leaving users responsible for enforcing its security boundaries.
What Has He Built?
Stanislaus Hees, founder of Teydra Lab, developed Autark to coordinate work that would otherwise require someone to move between separate AI assistants. Autark has a desktop application with a manager, specialist agents, shared tasks and memory.
The idea is that users give work to the manager, which then divides it between agents and brings their findings together. The aim is to retain responsibilities and unfinished work between conversations, thereby reducing the need for someone to keep explaining what should happen next.
Autark doesn't introduce a new AI model. Instead, Hees has built software that brings existing AI services together, organising their work and giving them access to tools on the user’s computer.
How Does The Team Keep Working?
Users can give Autark a task to carry out immediately or arrange for it to act later, such as preparing a briefing each morning. Agents can also respond to messages or check on work already under way, so users do not have to start a new conversation each time something needs doing. As its FAQ puts it, “Agents do not wait for a prompt.”
By default, Autark keeps running in the system tray after its window is closed, provided the computer stays on. This allows work to continue between visits, although users still need a clear account of what has been completed, what remains unfinished and whether anything needs their attention.
Which AI Services Does It Use?
Autark connects to ChatGPT plans through OpenAI Codex and paid Claude plans through Claude Code. It uses the providers’ official command-line tools, which allow software to send instructions and receive responses without someone typing into a chat window.
It can also use Gemini CLI, GitHub Copilot CLI, Kimi Code, API connections and local models. However, several of these connections have been built without a recorded live test, so users should check whether their preferred service has been demonstrated working before relying on it.
Autark itself is free and open source under the Apache 2.0 licence, but users must provide access to the AI services it uses. Subscription charges, usage limits and API fees still apply, and several agents working on one assignment may make many requests before producing a result.
What Can Agents Access?
Autark’s agents can use the permissions available to the person running the application. That can include reading and changing files, running commands and using connected services. There is no built-in sandbox to isolate their activity or enforced approval process to stop them taking particular actions.
During setup, the manager asks how much freedom it should have, but the project explains, “That is a question, not a technical barrier.” Choosing a cautious response gives the model instructions to follow but doesn't remove its ability to act.
For example, telling an agent not to send messages would not prevent it from using a mailbox it can access. To make that restriction dependable, a business would need to limit the account’s permissions or keep the mailbox disconnected.
Does Keeping Data On A PC Make It Private?
Autark says it stores chats, tasks and memory on the user’s computer, with encrypted traffic passing through its relay server between connected devices. However, an agent may still send information to the AI provider handling its requests, including material from private files it opens. Local storage therefore does not mean that everything stays on the PC.
Passwords and keys are protected in storage using operating-system encryption. Agents normally refer to them through identifiers, helping keep the actual values out of prompts. That reduces the chance of accidental disclosure, but a command given access to a credential can still use it inappropriately.
Businesses really will need to check both how credentials are protected and what they allow an agent to do, and they will also need to understand which documents or other information might reach an external AI service during the work.
How Much Has Been Proven?
Autark is still an early project, and its repository warns against assuming that publicity means every workflow will succeed. It offers a purchasing comparison using fictional information as a test, encouraging users to examine the tasks assigned to each agent and the results they produce.
That gives users something practical to try, but doesn't establish any consistent productivity gains in business use. Several agents might save someone the effort of coordinating a job, or leave them spending longer resolving conflicting answers and correcting mistakes. The useful comparison is really how much work remains before the result can be trusted.
What Does This Mean For Your Business?
For UK businesses interested in Autark, an initial trial should use public or fictional information in an isolated environment without business credentials. Choose a task whose answer can be checked, then compare the quality of the finished work, the time spent reviewing it and the AI charges with the existing approach.
Before extending the trial, establish which files, accounts and services the agents can reach, how their activity can be reviewed and how access can be withdrawn. Those limits should be enforced through account permissions and separate working environments, so they remain in place even if an agent fails to follow an instruction.
Hees’s project shows that building a team of AI agents is becoming more accessible. For businesses, the next question is whether that team makes a familiar job easier to complete. A worthwhile result would save time after checking and corrections, stay within an agreed budget and leave the business in control of its information and systems.