DeepSeek Super-agent: Open-Source Rival to Claude Code

DeepSeek has open-sourced a prototype coding agent called dsh-superagent, and it is aimed squarely at Anthropic's Claude Code. The DeepSeek Harness team announced it on WeChat on October 31 and put the code on GitHub under the MIT license. Unlike a chatbot-style assistant, it watches a repository (and optionally a Feishu chat) and starts work when an authorised person asks, without anyone babysitting a session.

Quick summary

• What it is: a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness.
• Licence and status: MIT, v0.1.0, labelled a research prototype and "not an official DeepSeek product" in its own README.
• How it works: a Router decides what counts as real work; isolated Worker sessions do each job and reply in the original conversation, often with a pull request.
• Claimed speed-up: in DeepSeek's simulation, p95 task time of 15.7 minutes versus 29.5 minutes for a single-session baseline.
• Catch: those numbers come from synthetic messages and simulated tools, and the agent runs with your OS user's full access.

What is DeepSeek's Super-agent?

According to the repository's README, dsh-superagent listens to GitHub and, if you enable it, Feishu (ByteDance's work chat app, called Lark outside China). When an allowed person asks it for something, it opens a work item, works on it in its own agent session and directory, and answers where the request came from. A local web panel lists only the items that need a human.

The guiding rule in the README is blunt: 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 speaks again only with a result, a question or a failure. Bloomberg's report, carried by The Straits Times, adds that it can follow developer group chats and suggest fixes without being summoned.

How does the Router-Worker design work?

MarkTechPost's breakdown of the repo's docs describes two layers. A long-running Router handles incoming events and decides whether a batch of them really asks for new work. Each accepted request becomes a Ticket with a recorded reason it may proceed, and an isolated Worker session executes it. Later comments on a pull request are routed back to the right Worker, so you do not have to mention the agent again.

Proactive behaviour is gated. The authorisation docs define three accept modes: off, observe and auto. Observe is the default, which means the agent only logs what it would have accepted until a team deliberately turns on auto.

Is it really faster than a single agent session?

DeepSeek tested the design against simulated group-chat and GitHub CI streams of roughly 300 messages, about 20 of them genuine work requests, with five runs per setup. MarkTechPost, citing a Chinese-language write-up of DeepSeek's design post, reports the following:

That is about 1.9x faster at the slow end. Two honest caveats: the figures are self-reported and secondhand, and they measure how quickly tasks were dispatched and finished under a noisy message stream, not how good the resulting code was. MarkTechPost notes that real coding quality was not measured. Treat the table as evidence for the architecture, not a verdict against Claude Code.

How does it compare with Claude Code and Copilot?

MarkTechPost's side-by-side puts the three approaches like this (the "not disclosed" cells are the outlet's wording):

Can you run it, and what does it cost?

There is no hosted service or price list. You run it yourself as a profile inside DeepSeek Harness, which is itself open source under MIT and in developer preview. The prototype needs Node.js 22.19 or newer (or 24+), git, Python 3.8+, tmux and the GitHub CLI, plus a model provider you configure. Because it simply uses whatever model your Harness profile points to, your running cost is whatever that provider charges. For Indian developers that means there is no subscription to buy, but you do need API access and a spare GitHub account for the agent.

Security deserves a real pause. MarkTechPost warns that the agent runs with full access of the operating-system user, so isolate it in a container or VM and start in observe mode. An early GitHub discussion thread from a tester also reported two bugs in the first hour of use, which fits the README's own warning that the code may contain bugs.

Why does this matter for the AI race?

The release lands as Chinese and US labs compete on cheaper, more autonomous agents. Bloomberg notes DeepSeek has raised roughly 100 billion yuan ahead of a planned IPO, and its Harness team was formed this year to build agents that rival US coding tools. If you follow the model side of the race, our explainer on Gemini 4 Carbon against Claude Opus 5.5 and GPT-6 covers where the frontier models stand, and PC owners weighing local agents can read about NVIDIA RTX Spark laptops.

Related articles

Gemini 4 Carbon: Access, Price and How It ComparesFortnite v43.00 Patch Notes: Gold Bars Vaulted, New MapIntel Nova Lake: 20% Gaming Leak, bLLC and Release DateWoW Forever Launch Time in India (IST), Price and Preload