DeepSeek Super-Agent: What the Claude Code Rival Actually Is

What DeepSeek released

The repository is deepseek-ai/dsh-superagent-prototype on GitHub, released under the MIT license. TechNode reported on November 2, 2026 that DeepSeek's Harness team had open-sourced it as a research prototype, and Digital Trends and Interesting Engineering covered it on November 1.

The README describes a proactive coding agent for a single GitHub repository, packaged as a bundle for DeepSeek Harness, the open-source agent harness DeepSeek released earlier in 2026. It listens to GitHub and, optionally, Feishu, the workplace chat app. When an authorized person asks for work, it opens a work item, does the work in its own agent session and directory, and replies in the original conversation, for example with a pull request. A web panel shows which items need a human.

The README opens with a plain notice: research prototype for reference only, not an official DeepSeek product. It adds that much of the code was written by coding agents, that the published version differs from the internal one, and that some designs in the accompanying article are future directions that are not implemented.

Why it gets called a Claude Code rival

The rival framing comes from context more than from the code. DeepSeek formed its Harness team earlier this year to build agents that compete with the coding products of US labs, and coverage has framed Super-Agent as a direct challenge to Anthropic's Claude Code. Digital Trends was more measured, calling it an early sign of where the industry is going and not a reason to uninstall Claude Code or Codex.

The comparison is also slightly off target. Claude Code is an interactive agent harness you drive from a terminal, IDE, or desktop app. DeepSeek's direct equivalent of that is DeepSeek Harness itself, which the existing comparison of DeepSeek Harness and Claude Code covers as an evaluation method. Super-Agent sits one layer up: it is a coordinator that decides which requests become work and runs harness sessions on its own. The closer comparison is a background agent that turns tickets or chat requests into pull requests.

What the coverage overstated

Two claims deserve correction before anyone acts on them.

It does not simply decide on its own. Several write-ups, and the video description, describe an agent that reads team chats and starts work unprompted. The README is narrower: it acts only when asked or when a rule the team wrote allows it, and it speaks only when someone must act or can use a result. Proactivity is a policy the team configures, not open-ended initiative. That is the correct design, and it is less dramatic than the headlines.

There is no quality benchmark. The figure in coverage is a 95th-percentile task time of 15.7 minutes for the super-agent design against 29.5 minutes for a single-session subagent baseline, with a median of 12.4 minutes. Those runs were simulated, on synthetic messages and simulated tools, and no coding-quality measure was reported. A latency percentile from simulation says something about coordination overhead and nothing about whether the output was correct. The README itself contains no benchmarks; the numbers come from the accompanying article.

The idea worth taking seriously

The README states the goal as taking as little human attention as possible while giving as much value as possible. That is the right target. Anthropic reported in October that people supervising agents are now the constraint, a point covered in humans as the bottleneck in managing AI agents. A person can only watch a few sessions at once, so throughput depends on how rarely the agent needs one.

Super-Agent's answer has three parts. Requests become discrete work items, each with its own session and directory, so failures stay contained. Results go back where the request came from, so nobody hunts for them. And a panel lists only the items that need a decision, so silence means nothing needs attention. Those patterns transfer to any harness, including Claude Code driven by a scheduler or a ticket queue, and they echo how long-running agents keep state outside the model.

The risks of an agent that listens to chat

An agent that turns chat messages into code changes widens the attack surface. Every message an authorized person can post, or paste, becomes potential instructions. The README routes work only from authorized requesters and documents authorization separately, which is the minimum. Teams should also treat quoted logs, issue bodies, and forwarded messages as untrusted input, since that is where injected instructions hide, a problem explained in where untrusted input enters a coding agent. Give the agent its own GitHub account with write access to one repository, require human review before merge, and keep secrets out of its environment.

Cost is the other quiet risk. A background agent that reproduces bugs and drafts fixes for every request makes many model calls, including ones that end in silence. Re-reading the same code, docs, and past decisions on every work item is where spend accumulates. Retrieval over an indexed copy of that material is the usual fix; on repeated or reference-heavy work, RDK reports 80 to 90 percent token reduction by retrieving answers instead of regenerating them.

How to evaluate it

Run it on a sandbox fork, not your main repository. Pin the DeepSeek Harness version the bundle expects. Give it ten real requests from your own history: bug reproductions, small fixes, and one ambiguous ask. Record for each whether it acted when it should, stayed silent when it should, and produced a pull request you would merge. Count the human minutes spent on each item, including review. Compare that against the same ten requests handled in an interactive harness. If Super-Agent does not reduce minutes per merged change, the design idea may still be worth copying while the prototype is not worth running.