AI: DeepSeek Develops Agents Capable of Working Almost Independently

DeepSeek wants its AI agents to complete more work with less human supervision. On October 31, 2026, the Chinese company’s Harness team unveiled a proposed architecture called Super-agent, designed to coordinate several tasks simultaneously. Its researchers say one person can usually supervise only three to five active AI agent sessions. The new system aims to overcome that limit by automating task distribution and coordination.

DeepSeek Wants AI Agents to Work with Less Supervision

DeepSeek has presented a new approach to managing autonomous AI agents. The proposal focuses on a problem that has become increasingly visible as companies deploy assistants capable of writing code and completing complex assignments.

DeepSeek has already strengthened its position in the global AI race with its V4 models. Now, its researchers are looking beyond the capabilities of individual models.

Their concern is human attention.

According to the DeepSeek Harness team, a person can generally keep track of three to five active agent sessions. Managing more requires switching between conversations, reviewing intermediate results and repeatedly clarifying instructions.

The Super-agent proposal changes how that work is organized.

Instead of having users supervise several independent assistants, DeepSeek introduces a central coordinator called a Router.

The Router receives requests, identifies what needs to be done and assigns each task to a separate Worker.

Workers operate in their own sessions, allowing several assignments to progress at once. They report their results through the coordination system rather than constantly interrupting the user.

A software developer could therefore request a bug fix, a code review and an investigation into a failed test without personally managing three separate conversations.

The system would distribute the requests and notify the developer when a decision or review is needed.

DeepSeek has also released an open-source prototype under the MIT license.

However, the company makes an important distinction. This is research software, not an official finished product. The public implementation also differs from the internal system described in its design proposal.

A System That Can Respond Without New Prompts

The architecture is designed to handle more than instructions entered into a chatbot.

DeepSeek wants its agents to respond to events generated by the tools people already use.

A GitHub comment, a software test failure or a message in a team discussion could become a new task.

The Router evaluates the event and determines whether a Worker should handle it.

In an example described by DeepSeek, an internal agent monitors a group conversation, identifies a reported software problem and prepares a proposed fix without being directly mentioned.

The objective is to avoid unnecessary interruptions.

Workers can also maintain the context of an assignment across different conversations. A task initiated in a group discussion could continue through comments on a GitHub pull request.

DeepSeek has published preliminary results from simulated workloads.

In those tests, its Super-agent system completed half of the requests in 12.4 minutes or less, compared with 20.4 minutes for a conventional single-session configuration.

The experiment measured coordination performance under simulated conditions. It did not independently establish that the agents write better code.

The architecture also consumes additional computing resources because several sessions can operate simultaneously.

Other technology companies are working on related approaches. OpenAI and Anthropic are developing tools that allow agents to use software and execute increasingly complex workflows.

The sector has also been examining the security vulnerabilities that autonomous AI agents can introduce.

For DeepSeek, the next step involves demonstrating that improved coordination can operate reliably outside controlled tests.

Greater Autonomy Still Requires Human Control

DeepSeek’s proposal does not remove people from the process.

The published prototype includes authorization rules determining who can assign work and under what conditions an agent may act.

Certain operations still require human approval.

The system is primarily designed for software development, where agents can examine repositories, suggest changes and prepare pull requests.

That environment creates practical security concerns.

An autonomous assistant may encounter malicious instructions hidden inside documents, code comments or external messages. If it has excessive pe