API
< Back to all clusters
[TECHNOLOGY] · China · 22 sources

started · updated

DeepSeek releases Super-agent coding prototype to rival Anthropic

DeepSeek has released a research prototype for a new AI coding architecture called “Super-agent,” designed to reduce the human attention required for multi-task workflows. Released under an MIT license by the company’s Harness team, the system aims to move beyond passive chatbots toward autonomous collaborators that can proactively monitor environments.

The architecture utilizes a two-layer system consisting of a Router and Worker sessions. The Router remains constantly active to receive and distribute tasks from sources such as GitHub, Feishu group chats, and emails, while individual Workers execute specific tasks like code reviews or error reproduction in isolated sessions. A Watchdog mechanism is also included to monitor task progress and detect errors.

In testing, the system reportedly reduced task completion times from approximately 29 minutes to nearly 16 minutes. While the prototype can generate pull requests and respond to developer discussions without being explicitly prompted, DeepSeek notes it is research software and not an official, production-ready product. This development positions DeepSeek in direct competition with tools like Anthropic’s Claude Code and OpenAI’s Codex.

Entities

Anthropic · DeepSeek · DeepSeek Harness · Feishu · GitHub · OpenAI

Claims

What the coverage asserts, and how many sources carry each claim.

Sources