DeepSeek previews super agent for coordinating coding assistants

DeepSeek has previewed a plan for a "super agent" that coordinates multiple AI agents with minimal human intervention, in a technical article from its DeepSeek Harness team published on Saturday, the South China Morning Post reports.

"Single-task limits depend on model capabilities, while multitask limits depend on human attention," the paper said. The team estimated that a person can typically oversee only three to five active AI agent sessions at once.

Under the proposed architecture, a central coordinating agent would distribute assignments among independent agents working in parallel, which report their progress back to it. Agents could also respond to events such as messages and code updates without explicit instructions. In one example, an agent could detect a bug reported in a group chat, try to reproduce and fix it, and submit a proposed code change for human review, or stay silent if it found no viable fix.

In experimental tests, the approach completed one task in just under 16 minutes, compared with more than 29 minutes for a conventional approach. DeepSeek acknowledged security and privacy risks, including malicious instructions compromising agents connected to other services. The proposal comes as Anthropic's Claude Code and OpenAI's Codex take on more complex software tasks.