The DeepSeek super-agent is a coding agent that waits to be asked, then runs many jobs in parallel. DeepSeek’s Harness team released it as an MIT-licensed prototype called dsh-superagent. The headline claim is a 1.9x faster p95 task time, but that number comes from a simulation. This guide explains what it is and what it does not prove.
Key takeaways
- What: dsh-superagent, a proactive GitHub coding agent from DeepSeek’s Harness team.
- Status: research prototype, v0.1.0, MIT licence.
- Claim: simulated p95 task time of 15.7 minutes vs 29.5 for a single-session agent.
- Catch: tested on synthetic messages and simulated tools. Real code quality was not measured.
- Safety: it runs with full access to your OS user. Isolate it.
Last updated: Monday 2 November 2026, IST. Status: prototype, self-reported results.
[Insert Rank Math Table of Contents block here]
What is the DeepSeek super-agent?
The DeepSeek super-agent is a proactive coding agent for one GitHub repository, according to MarkTechPost. It listens to GitHub and, optionally, Feishu. When an allowed person addresses it, it opens a work item called a Ticket. Each Ticket runs in its own session and folder, and results often come back as a pull request.
Its stated rule is to take as little human attention as possible while giving as much value as possible. In practice, it confirms once, stays quiet while working, and writes again only with a result, a question or a failure.
Why DeepSeek built a super-agent
DeepSeek argues that most coding agents treat one session as the basic unit. Per MarkTechPost’s account of the report, a person can actively follow only three to five sessions at once, so human attention, not model skill, limits multitasking. The DeepSeek super-agent is built around that limit.
How the Router-Worker design works
There are two layers. A long-running Router handles messages and dispatch. Isolated Worker sessions do the actual work.
- Trigger: an event from GitHub or Feishu.
- Eval: decides if a batch of triggers asks for new work.
- Ticket: an authorised unit of work with a fixed record of why it may proceed.
- Subscription: later pull-request comments reach the right Worker without a new mention.
Proactive work is gated by three modes: off, observe (the default) and auto. In observe mode it only records what it would have accepted. That default is a sensible safety choice.
The DeepSeek super-agent speed claim, checked
DeepSeek tested with simulated group-chat and GitHub CI message streams. Each run had about 300 messages, around 20 of them real work requests, and each setup ran five times. All figures are self-reported and reach us through secondary reports of DeepSeek’s post.
| Setup | p50 | p95 |
|---|---|---|
| DeepSeek super-agent | 12.4 min | 15.7 min |
| Single session plus subagents (neutral prompt) | not given | 29.5 min |
29.5 ÷ 15.7 is about 1.9x. But the baseline called a subagent zero times in five runs under the neutral prompt, so it was effectively working one task at a time. That makes the comparison favourable by design. Treat it as a demonstration of the idea, not proof that the DeepSeek super-agent is better at coding.
As message arrivals rose from about one to 40 per minute, batch size grew from 1.0 to 2.8 and per-request latency rose from 2.5 to 3.9 seconds. Removing direct event pass-through raised router decisions by about 30%.
How it compares with other coding agents
| DeepSeek super-agent | Claude Code Action | Codex | |
|---|---|---|---|
| Trigger | Mention by an allowed person | @claude mention, issue, workflow | @Codex in Slack; cloud tasks |
| Runs where | Your machine | Your GitHub Actions runner | Codex cloud |
| Licence | MIT | MIT | CLI agent is open source |
| Status | Research prototype | Available | ChatGPT Plus and above |
This table follows MarkTechPost’s comparison; fields marked “not disclosed” there are left out. Verify the others on each vendor’s page.
What the DeepSeek super-agent needs and risks
- Node.js 22.19+ or 24+, git 2.32+, Python 3.8+, tmux and GitHub CLI 2.48+.
- It runs with the full permissions of your OS user, per the report. Use a separate machine or container.
- It is one commit at v0.1.0. Expect rough edges and missing docs.
What the DeepSeek super-agent means for India
Indian developers and startups get a free, MIT-licensed DeepSeek coding agent they can read and modify, with no subscription. The cost is setup effort and your own model bill. It is worth a weekend test if you already use GitHub heavily; it is not ready to replace a paid agent.
What we don’t know yet
- Real-world code quality and bug rates.
- Which model it uses by default; it follows your Harness profile.
- Whether DeepSeek will ship a hosted version.
- Independent tests.
Watch: DeepSeek super-agent walkthrough
[Hari: embed a video or add your own install screenshots.]
Frequently asked questions
Is the DeepSeek super-agent free?
The software is MIT-licensed. You still pay for whatever model your Harness profile uses.
Is it better than Claude Code?
Unknown. DeepSeek’s test compared it with a single-session baseline in a simulation, not with Claude Code on real projects.
Is the DeepSeek super-agent safe to run?
It runs with full access to your OS user, so isolate it. Its default mode only observes.
What is p95?
p95 is the time within which 95% of tasks finish. It shows slow cases, not the average.
Hari’s take
The idea is good: stop counting sessions and start counting your attention. The evidence is thin. A simulation where the rival never uses subagents mainly shows the design works as designed.
I would test this Claude Code alternative, the DeepSeek super-agent, on a throwaway repo and publish my own numbers. Until then, nobody should write that it beats anything. For Anthropic’s side of the race, see our Claude Mythos 5.5 rumours post.
Sources: MarkTechPost, 1 Nov 2026 · Bloomberg headline, 1 Nov 2026. Last updated 2 November 2026.
