DeepSeek Releases a Coding Assistant That Needs No Chasing
DeepSeek's Harness team open-sourced dsh-superagent-prototype on October 31: a coding-assistant prototype that lives in a repository, listens for requests on GitHub and Feishu, starts work on its own and replies with pull requests. It is MIT-licensed, and the team states plainly that it is a research prototype.
- 01THE RELEASE — The Harness team published the post and open-sourced the prototype on October 31.
- 02THE ROUTER — The routing layer monitors the repository and Feishu, identifying approved requests.
- 03THE WORKER — Each ticket gets its own session and folder, and work starts on its own.
WHAT HAPPENEDWhat Happened
DeepSeek's Harness team published a post on WeChat on October 31 and open-sourced dsh-superagent-prototype, a coding-assistant prototype that resides permanently in a team repository, under the MIT license. It does not wait to be asked: it monitors a GitHub repository and an optional Feishu group chat, and once an authorized member sends a request, it starts work on its own.
The design has two layers. A resident routing layer reads requests and opens a ticket for each approved one; every ticket gets its own work session and folder, and when the work is done it returns to the original conversation to deliver, most typically a pull request. Running it requires a dedicated GitHub account, Node.js 22.19 or later and GitHub CLI 2.48 or later, and it supports github.com only.
In the team's own synthetic tests, across roughly 300 messages and about 20 real requests, the super-assistant's median completion time was 12.4 minutes, with a 95th percentile of 15.7 minutes, against 29.5 minutes for a baseline of a single session with sub-assistants. The cost was far higher cached-token usage: 23.2 million per run, versus 6.1 million to 10.8 million for the baseline.
ATTENTION BUDGETWhat It Saves Is Attention, Not Tokens
DeepSeek's design starts from a practical fact: a person can only keep track of 3 to 5 active assistant sessions at a time. The bottleneck is not how fast a model writes code, but the bandwidth of human checking. The super-assistant's answer is to collapse checkpoints from process to result: review the pull request at delivery.
The trade-off has a price. Token usage multiplies in exchange for humans not having to watch throughout. When attention costs more than compute, the math works. When compute costs more than attention, the math needs redoing.
RESEARCH LABELResearch Prototype: a Disclaimer and an Honest Label
The repository describes itself bluntly: a research prototype for reference only, not a formal product; the code may contain bugs, and much of the engineering was written by a coding assistant itself. That is both a disclaimer and honesty. Calling an unfinished artifact unfinished is more credible than packaging it as a product.
For users, that means reviewing everything before deployment. Authorization scope, repository permissions and the Feishu integration must each be reset against one's own security baseline; the default configuration should not be copied as-is.
THE BASELINEThe Baseline Reveals the Real Opponent
The baseline is the interesting part. DeepSeek did not compare against human engineers, but against the conventional setup of a single session plus sub-assistants. What the super-assistant beat is another way of using assistants, not people.
That shows the competition has changed. The coding-assistant arena is shifting from model capability to orchestration. Whoever manages multiple sessions well can deliver faster on the same model.
THE SHIFTWhat Changed
SIGNAL JUDGMENT — ANALYSISAssistant Competition Shifts to Orchestration
Tokenta Signal's judgment: what DeepSeek is selling this time is not a model, but a way of working. Having people watch less of the process and verify more of the result, orchestration is becoming the real moat of coding assistants.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
The original confirms that DeepSeek's Harness team open-sourced dsh-superagent-prototype on October 31 under the MIT license; it resides in a repository, monitors GitHub and Feishu requests and delivers pull requests, and the repository describes itself as a research prototype, not a formal product.
PRIMARY SOURCESPrimary Sources
- GitHub: deepseek-ai/dsh-superagent-prototype
https://github.com/deepseek-ai/dsh-superagent-prototype - AI Affairs: DeepSeek Harness team releases coding agent prototype for GitHub
https://www.aiaffairs.com/technology/deepseek-harness-super-agent-coding-prototype/
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook