The DeepSeek Harness team announced the Super-agent prototype, an AI that runs continuously, accepting jobs from the chat group and GitHub, with a reminder that this is just a test.

On the evening of October 31 (Beijing time), DeepSeek's DeepSeek Harness team announced the Super-agent architecture and opened the source code of an accompanying prototype. The idea is that an AI runs continuously, monitoring the programmer's chat group and the code repository on GitHub, then automatically detects errors even though no one mentions it or assigns work. "Harness" is software surrounding the language model, helping AI handle multi-step tasks.

DeepSeek says an internal run did just that. When someone reports an error in the chat group, even without calling AI, the system silently tries to reproduce the error, tries to fix it, and then sends a pull request to the reviewer. Only when it confirms the error and has a simple fix does it speak up, otherwise it remains silent. In some cases, the group is still discussing how to handle the problem when a correction has already been submitted.

The problem DeepSeek raises is human attention. The AI ​​does a job that requires less and less intervention, but according to the authors, a person can only monitor about 3 to 5 sessions at a time. The bottleneck thus shifts from the model to the user. The accompanying documentation aims to consume as little human attention as possible, while still providing as much value as possible.

To do this, the system is divided into two floors. The coordination layer is always running, only receiving messages and assigning tasks, not doing specific tasks itself. The underlying "executors" each have their own session and directory, running in parallel. When there are new comments in the edit, the message is transferred directly to the correct staff member, so users do not have to reopen the conversation from the beginning. The message must also pass through the filter: the code checks the sending source and the allowed list, then goes to the model for evaluation. In addition, there is a separate monitoring AI, watching to see if other AIs get stuck or go astray. In group chat, each reply is limited to less than 100 words to avoid disturbance.

DeepSeek has a test run. About 300 messages were included, of which about 20 were business requests. Half of the work is done in 12.4 minutes and 95% is done in 15.7 minutes. With a single session with the assistant calling tool, the 95% group took 29.5 minutes using neutral prompts, and 5 runs did not call the secondary assistant. In return, the Super-agent uses more tokens. The authors themselves note that the test is based on simulated load and a simplified environment, the results are for reference only.

The prototype currently targets a GitHub repository, possibly appending Feishu, ByteDance's messaging platform, and a portion of the design is yet to be implemented. DeepSeek emphasizes that this is not an official product. The authors also list four open issues: scaling, how humans intervene midway, coordination between multiple AIs, and security. As for security, they raise the risk when cloud AI accesses login information on users' computers, because it can be attacked by command injection (prompt injection). The code repository's permissions documentation describes the default mode as "observe": The AI ​​only records what it will do without taking action, and permissions are exercised by the system, not determined by the message content.

DeepSeek's article also admits that separating the "brain" from the "hands" is similar to the way Anthropic once announced, and that dividing the work between multiple AI subs is not a new idea. Where DeepSeek claims to be new is in incorporating familiar mechanisms of computer systems, such as early filtering, background running and monitoring, into an always-on AI. Observers consider this a step to challenge Anthropic's Claude Code.

Super-agent runs on DeepSeek Harness, the open source framework DeepSeek launched on August 13 under the MIT license and surpassed 140,000 GitHub stars in 4 days. The Super-agent warehouse was established on October 31, reaching 1,262 stars after 7 hours and more than 7,000 stars after two days.

Behind is a large capital source. According to information reported on October 30, DeepSeek's second round of capital raising reached 105.5 billion yuan, about 15.8 billion USD, equivalent to about 416 trillion VND, valued at about 615.5 billion yuan, or about 2.4 million billion VND. Nearly 30 investors participated, Tencent and CATL are still the two largest parties, state capital accounts for about 27%. DeepSeek is said to be expected to list in 2027.

Super-agent source code uses MIT license, written in TypeScript, version 0.1.0. It runs as a private profile inside DeepSeek Harness and only receives GitHub requests from authorized users.