DeepSeek Super-agent prototype is the most practical free coding tool to land this weekend. On October 31, 2026, DeepSeek’s Harness team published an open-source prototype that watches for coding requests, spins up a separate session for each one, and comes back with a pull request for the team to review.
A pull request is a proposal, not a merge: code waits for a human before landing.
What the prototype actually does
The public release is called dsh-superagent-prototype, and it is built for one GitHub repository at a time. It listens to GitHub events, and it can optionally connect to Feishu, a workplace messaging service, so requests can arrive from a team chat as well as from GitHub itself.
The design goal, in the team’s own words, is an agent that takes as little human attention as possible while giving as much value as possible. In practice that means:
- It stays available. Instead of waiting to be prompted, it receives messages and delegates jobs on its own.
- It separates the work. Each work item gets its own agent session and its own directory, so jobs do not tangle.
- It returns a proposal. Finished work arrives as a pull request the team can review, change, or reject.
DeepSeek is explicit: this is a research prototype, not an official product or production-ready software. It may contain bugs, differs from its internal version, and does not implement every idea in its design article.
The speed claim — and why it comes with an asterisk
The headline numbers come from a preliminary experiment the team ran on simulated messages, not on real client work. About 300 simulated messages carried roughly 20 work requests through the system.
In that experiment, the Super-agent approach completed half the tasks within 12.4 minutes and 95% within 15.7 minutes. A conventional single-session setup took 29.5 minutes to reach the same 95% mark.
Read that carefully. The comparison used simulated message streams and simulated slow tools, rather than measuring real-world code quality directly. The time gap is real inside the experiment, but it is not proof that your next feature ships twice as fast.
What the experiment does suggest is where the time goes in normal agent work: waiting, context-switching, and one long session juggling many jobs. Splitting jobs into separate sessions attacks exactly that overhead.
The evidence: what is confirmed and what is not
| Detail | What the source says |
|---|---|
| Published | October 31, 2026, by DeepSeek’s Harness team |
| Package name | dsh-superagent-prototype, for one GitHub repository |
| Licence | MIT — use, copy, modify and sell, with the licence notice included |
| Inputs | GitHub events, plus optional Feishu workplace messages |
| Output | A pull request per work item, for human review |
| Experiment scale | About 300 simulated messages, roughly 20 work requests |
| Team-reported times | 50% in 12.4 min, 95% in 15.7 min, vs 29.5 min for a single session |
| Experiment caveat | Simulated streams and tools; real-world code quality not directly measured |
| Status | Research prototype, not an official DeepSeek product; may contain bugs |
| DeepSeek V5? | No. Separate test-file names are not a release, and DeepSeek has announced no V5 model |
A pull request is a formal proposal to merge code changes into a project. Teammates review the diff, comment, and approve or reject it. Nothing merges by itself.
How a request flows through the Super-agent prototype
1. Request arrives — A GitHub event or a Feishu message describes a coding job.
2. Session spins up — The job gets its own agent session and working directory.
3. Agent works alone — It codes without blocking other jobs or the team’s chat.
4. Pull request returns — Finished work comes back as a reviewable proposal.
5. Human decides — A developer reviews, requests changes, or merges.
How to profit from it
Free tooling does not kill a service business. It moves the billable part from typing code to owning the queue, the review, and the result.
- Sell a “request queue” service. Offer small teams a managed intake: they file GitHub issues, you run the prototype, review every pull request, and merge only what passes your checks.
- Charge for the review, not the generation. The prototype writes the first draft for free. Your fee is the code review, the tests, and the accountability when something breaks.
- Start with one repository. The prototype is scoped to a single repo. Pick a client’s smallest repo, prove the workflow, then expand.
- Keep it sandboxed. An agent that runs code needs an isolated environment. Never point a prototype at a production database or a live server.
- Do not promise the 15.7-minute number. That figure comes from a simulated experiment. Quote your own measured turnaround after a trial week.
Sources
DeepSeek’s Harness team release as reported by Business Today, Daily Inference and Asia Daily, citing the team’s WeChat post and GitHub manual. Check the prototype repository itself before deploying anything.
Frequently Asked Questions
Is DeepSeek’s Super-agent a finished product?
No. DeepSeek calls it a research prototype, not an official product, and warns it may contain bugs and differs from its internal version.
What does it cost?
The prototype is MIT-licensed open source, so there is no licence fee. You still pay for your own compute, hosting, and the model calls behind it.
How many repositories can one prototype handle?
One GitHub repository per installation, according to the public prototype description.
Does it merge code automatically?
No. Work returns as a pull request, which a human reviews before anything merges.
Are the speed numbers from real client work?
No. They come from a preliminary experiment with simulated messages and simulated tools, so treat them as a directional signal, not a guarantee.
Is this the same as DeepSeek V5?
No. DeepSeek has announced no V5 model. Names spotted in test files are not a release.
Keep Reading
Want more tools that actually pay for themselves? Browse the AI Tools section for the latest launches, pricing breakdowns, and profit playbooks — and check the official source linked above before you buy or deploy anything.