BEIJING (Realist English). On 31 October, the DeepSeek Harness team published an announcement on WeChat and posted on GitHub a prototype of an autonomous coding agent called Super-agent. The code is distributed under the MIT license, and the project itself is positioned as a research prototype rather than an official company product. In the first 24 hours after publication, the dsh-superagent-prototype repository gained more than 4,000 stars.
What Super-agent can do
The key difference between the new tool and existing solutions is proactivity. The agent constantly works in the background, monitoring one GitHub repository and, optionally, group chats in Feishu (Lark). When an authorized team member requests work, the agent opens a work item, performs the task in its own session and working directory, and returns the result to the original conversation — for example, as a pull request.
The architecture is based on a two-tier system. A constantly active Router receives all human requests, maintains shared context and distributes tasks, but does not perform them itself. Isolated Worker sessions run tasks in parallel with a full set of tools and report back to the Router without communicating with the user directly. Each Worker subscribes to the branches it is interested in (for example, a pull request or a chat topic), so subsequent events go directly to it, bypassing the Router. An external Watchdog process ensures that Workers do not hang and make real progress.
The system is event-driven: chat messages, review comments and CI results serve as triggers. Filtering happens in two stages — first, deterministic code filters out most of the noise, then a language model evaluates the remaining candidates.
Performance and limitations
In preliminary tests on a synthetic workload (about 300 messages and roughly 20 work requests), Super-agent completed half of the tasks in 12.4 minutes and 95% in 15.7 minutes. For comparison, a single-seat scheme with a subagent tool showed a p95 of 29.5 minutes. The team notes that these are only preliminary results and calls them a benchmark, not a full-fledged one.
Among the open problems, the developers list scaling the Router, long-term memory, inter-agent interaction and the risk of prompt injections, since the cloud agent has access to credentials on the user’s local devices. Deployment requires a dedicated GitHub account for the agent, Node.js 22.19 or newer, GitHub CLI 2.48 or newer, and a configured model provider in DeepSeek Harness. Only github.com is supported, not GitHub Enterprise Server.
Agent race and IPO preparation
The release of Super-agent is part of DeepSeek’s broader strategy to enter the market of autonomous agents for developers. The company created the Harness team in early 2026 specifically to compete with Anthropic, OpenAI and other American laboratories in the agent segment, not just chatbots. Earlier, in August 2026, DeepSeek had already released the open-source Harness agent, which was seen as a direct competitor to Claude Code.
The financial context adds weight to the announcement: in its latest funding round, DeepSeek raised about 100 billion yuan ($15 billion) — one of the largest rounds in Chinese history — ahead of an IPO scheduled for 2027. At the same time, the US has banned the use of DeepSeek in government systems, citing national security concerns.
Super-agent is still a research prototype rather than a finished product, and its real impact on the market will depend on how quickly the team solves the identified open problems — from scaling the Router to security when accessing local devices.
However, the vector is clear: DeepSeek is shifting competition from the plane of “which model writes code better” to the plane of “which agent better organizes team work with minimal human involvement.” This is precisely the segment where the main battle between Chinese and American laboratories is now unfolding, and the release of Super-agent into open access gives developers the opportunity not only to study a competitor’s approach but also to influence what the next generation of coding automation tools will look like.
