DeepSeek wants to reduce the dependence of its AI agents on human supervision. On October 31, 2026, the Chinese company presented an architecture called “Super-agent”, designed to execute several tasks simultaneously without constantly requesting the user. According to its researchers, a person can currently follow only three to five AI agent sessions at a time. DeepSeek proposes to overcome this limit thanks to an automated coordination system.
In brief
- DeepSeek presents an architecture that allows AI agents to work with less human intervention.
- The system distributes tasks between several agents and can react automatically to certain events.
- An open source prototype exists, but several functions remain experimental.
DeepSeek wants AI agents that can work alone
The DeepSeek Harness team has published a new technical proposal dedicated to the autonomy of AI agents. The objective is to limit the interventions necessary to supervise several assistants working simultaneously.
The company had already strengthened its competition with OpenAI thanks to new AI models. She is now interested in how these models can work together. The problem identified concerns human attention.
Today, a person can typically monitor between three and five active AI agent sessions. Beyond that, she must constantly switch from one window to another, checking the results and recalling the instructions.
DeepSeek wants to simplify this organization. Its architecture, called Super-agent, is based on a central coordinator responsible for receiving requests and distributing work to several specialized agents.
Each agent has its own session and can progress on a task without interrupting the others. A person could thus ask the AI to correct software, analyze a document and verify code. The system would distribute operations before transmitting results requiring validation.
The DeepSeek teams also published a prototype under the MIT license on GitHub. The code remains experimental. The company clarifies that it does not constitute an official product ready for use in production.
A system capable of reacting without waiting for an instruction
The novelty also concerns the way in which agents begin their work. In a typical configuration, the user gives an instruction, waits for a result, then decides the next step.
DeepSeek envisions more continuous operation. Its architecture includes a “Router”, responsible for communicating with users, and several “Workers”, responsible for executing tasks.
The coordinator can receive messages, GitHub comments, software test results, or notifications. It then decides which session should process each request.
Some events could even trigger an intervention without further human instruction. DeepSeek describes in particular an agent capable of spotting a problem reported in a collective discussion, searching for its origin and proposing a correction.
In this example, no one needs to mention the agent directly. The system can also preserve the context of a job between several exchanges. A correction requested in an email could thus be continued on GitHub without repeating the explanations from the beginning.
The first internal tests show encouraging results. On simulated tasks, the architecture completed half of requests in 12.4 minutes or less, compared to 20.4 minutes for a typical single-session setup.
This performance comes from DeepSeek testing and has not yet been independently confirmed. The coordination of several agents also leads to additional IT consumption.
Other companies are working on similar approaches. OpenAI and Anthropic are notably developing tools intended to automate complex tasks. Tremplin.io had already presented OpenAI tools intended for developers of autonomous agents.
The autonomy of AI agents still poses security problems
DeepSeek does not plan to completely remove human supervision. Users should retain the ability to control important decisions, including when an agent modifies software or accesses sensitive resources.
The prototype published by the company also contains several restrictions. Agents operate in separate environments and their actions depend on the permissions granted by users. Some functions described in the technical proposal are not yet available in this version. Security issues remain important.
An agent connected to messaging, code repositories, and external services can receive malicious information. It could also perform a wrong operation if its permissions are too broad.
These risks increase when several agents interact without permanent control. For businesses, the interest remains concrete. IT teams could automate more checks, fixes and repetitive tasks.
Crypto application developers could also be interested in this architecture to coordinate blockchain monitoring or analysis tools. However, no specific integration with a crypto protocol was announced in this presentation.
DeepSeek is therefore continuing its experiments on an AI capable of acting for longer and managing multiple requests without continuous assistance. The next difficulty will be to guarantee the reliability of these operations. Tremplin.io had previously documented the risks associated with autonomous AI agents when operating over long periods of time. For DeepSeek, reducing human surveillance will not be enough. It will also be necessary to demonstrate that its agents know how to respect the limits imposed on them.
Maximize your Tremplin.io experience with our ‘Read to Earn’ program! For every article you read, earn points and access exclusive rewards. Sign up now and start earning benefits.
