DeepSeek Shares Super-agent Design for More Autonomous Agents

As AI models need less human help on single tasks, the real bottleneck is shifting to human attention. Most current agents work in one session at a time. A person can usually follow only 3 to 5 active sessions, and juggling many tasks often leads to confusion and delays.

DeepSeek has shared its Super-agent design, aiming to turn agents into always-on collaborators that can handle many tasks in parallel. The system uses a two-layer setup with a Router and Workers, plus event triggers and a Watchdog to catch issues.

In early tests with about 300 messages, Super-agent finished tasks with a p95 time of 15.7 minutes, versus 29.5 minutes for the single-session setup. The team has released a prototype for testing, but notes that long-term memory and security issues still need work.