For software teams, DeepSeek’s proposed super-agent system is designed to reduce the human work of coordinating and monitoring multiple AI coding agents.
DeepSeek announced on 31 October that its Harness team had built the system as it competes with US rivals OpenAI and Anthropic in coding assistants. The company released it as an open-source research prototype on GitHub, while describing the work as an architectural exploration rather than a mature product.
At DeepSeek, the intended workflow starts in group chats. When team members report a software issue, the system can spot a possible defect and attempt to reproduce and fix it in the background, even when nobody has directly tagged it.
An internal instance has been running continuously, DeepSeek said, reproducing and fixing bugs, replying to code reviews and helping develop DeepSeek Harness. DeepSeek Harness is a plugin-oriented project distributed publicly on GitHub.
The design places a central coordinating agent over several independent agents. It would hand out assignments, allow them to execute work in parallel and collect their progress reports. The system uses a two-layer architecture, event-driven workflows, asynchronous multichannel interactions, separate storage and computation, and an external watchdog.
That structure is intended to let agents respond on their own to messages and code updates, without explicit instructions for each task. DeepSeek estimated that a person can typically oversee only three to five active AI-agent sessions at once.
In experimental tests, DeepSeek said the approach completed one task in just under 16 minutes, compared with more than 29 minutes for a conventional approach. Liu Ming, a Beijing-based AI entrepreneur, said the central challenge in AI software development is reliability, ahead of completing tasks faster.
Under DeepSeek’s proposal, an agent that finds a viable solution could submit a suggested code change for human review. Agents without a useful answer would stay silent, a safeguard intended to limit unnecessary interruptions.
The company’s move comes as Anthropic’s Claude Code and OpenAI’s Codex become increasingly capable of taking on complex software-development tasks. The contest is focused on coding agents that can handle such work with minimal human oversight.
DeepSeek also flagged security and privacy risks. Malicious instructions could compromise agents connected to other services and systems, it said.
Whether the system will become a public service remains unclear. If it is offered publicly in China, it could fall under the country’s generative AI rules, which include requirements on safety, data sources and content controls.
Comments (0)