DeepSeek open-sources Super-agent, a coding agent that runs parallel jobs from team chats
DeepSeek's Harness team released an MIT-licensed research prototype of a coding agent that turns GitHub and Feishu requests into parallel jobs, with a router layer meant to spare people from juggling agent sessions.
DeepSeek has published the code for a coding agent that is meant to sit inside a software team's GitHub repository and chat, take on work when someone asks, and run several jobs at once without a person juggling the sessions. The Chinese AI lab's DeepSeek Harness team released the project, dsh-superagent-prototype, under the MIT license on Saturday, alongside a technical article titled "Exploring Super-agent Design: Towards a More Autonomous Agent System." By Sunday afternoon the repository had more than 5,700 stars on GitHub.
DeepSeek is careful about what it is shipping. The README calls the code a research prototype "for reference only, not an official DeepSeek product," says much of the engineering was done by coding agents, and warns that the public version differs from the one the team runs internally. Some ideas in the article are not built into the release at all.
The South China Morning Post framed the release as part of DeepSeek's push against OpenAI and Anthropic in coding assistants, and The Next Web, citing Bloomberg, reported that the agent is aimed at Anthropic's Claude Code.
The problem it is built around
The team's argument is that the bottleneck in agent work has moved from the model to the person supervising it. According to IT Home's account of the article, DeepSeek observed that one person can usually follow only three to five active agent sessions at a time. Beyond that, work stalls while a human switches between windows, copies context from one session to another, or waits for a long task to finish before giving the next instruction.
The project's philosophy document turns that into a single design rule: take as little human attention as possible, and give as much value as possible. In practice the agent confirms a request once, stays quiet while it works, and writes again only when it has a result, needs a decision, or cannot finish.
How the router and workers split the job
The design has two layers. A long-running Router receives everything that comes in, decides whether it is a request for work, and keeps the overall picture. Each accepted job becomes a ticket that runs in its own isolated Worker session and working directory, so a slow bug fix does not block a code review that arrives a minute later.
In the open-source release, the agent watches one GitHub repository and, optionally, chats in Feishu, ByteDance's workplace messaging app. A ticket subscribes to the issue, pull request or chat thread it came from, so later messages there, such as review comments, can continue the same job in the same session. The execution document describes a detail that matters for responsiveness: shell commands that run longer than five seconds are pushed into the background, so a worker can read new messages while a build or test suite is still running.
Unprompted work is gated. The authorization document defines three modes for work nobody explicitly asked for: off, observe and auto. The default, observe, only records what the agent would have accepted, and the team recommends switching to auto only after reviewing those records. Explicit requests from people on an allowed list work in every mode.
The agent runs on the operator's own machine as a DeepSeek Harness profile. The setup guide lists Node.js, git, tmux, Python and the GitHub command-line tool as prerequisites, and the execution document notes that worker sessions run commands without approval as the operator's user account, which is a reason to isolate it before trying it.
What DeepSeek's own test shows
DeepSeek tested the design on a simulated team: streams of group chat and GitHub CI messages, each run containing about 300 messages of which about 20 were real work requests, with each setup run five times. Super-agent finished tasks with a median time of 12.4 minutes and a 95th-percentile time of 15.7 minutes, according to IT Home and MarkTechPost. A conventional single session with a subagent tool, given a neutral prompt, reached 29.5 minutes at the 95th percentile and, MarkTechPost reported, did not call its subagent once in five runs.
The figures are self-reported and come from a synthetic workload. DeepSeek says the experiment measured scheduling and coordination, not how well the agent actually writes code. The test also showed why the subscription shortcut matters: removing it, so every follow-up passed through the Router, raised the Router's decisions by about 30 percent.
Where it fits
The release lands in a crowded month for open models and agents. Mistral released open weights for Large 4, a 1-trillion-parameter model aimed at coding and cyber defense, Moonshot said Kimi K3.1 Pro's open weights are due by November 4, and Anthropic's IPO filing, from the maker of Claude Code, showed $16.3 billion in first-half revenue. DeepSeek's contribution this time is not a model but an architecture that works with whatever model the Harness profile is configured to use.
According to IT Home, the article also lists open problems the team has not solved, including how the Router scales as tasks grow, long-term memory, communication between agents, and security risks such as prompt injection and credentials leaking from the operator's machine. DeepSeek says an internal Super-agent instance built on the same ideas is already handling bug reproduction, code-review responses and work on DeepSeek Harness itself.
