CN — TechNode

DeepSeek team open-sources proactive coding agent research prototype

AINovember 2, 20261 min read

DeepSeek's Harness team has released dsh-superagent, a research prototype that takes coding requests from GitHub and Feishu, works on them, and sends results back. The code is open source under the MIT license.

What happened

DeepSeek's Harness team published an open-source project called dsh-superagent. It is a coding helper that receives development requests through GitHub and, optionally, Feishu (a team chat and work app). It works on each request in its own separate session. When it finishes, it sends results, such as a pull request (a proposed code change for review), back to the original conversation.

Key facts

It is set up around one GitHub repository.
It runs as a bundle for DeepSeek Harness.
Each task gets its own working folder and agent session.
It includes a web panel that shows items needing a person's attention.
It is released under the MIT license, a permissive open-source license.

How it works

The system turns approved requests into work items. Each item gets a separate space to work, so tasks do not mix together. A web panel lets people see which jobs are waiting for them. The project came from the team's own daily development work.

Why it matters

The project tests whether AI coding agents can handle longer jobs with less help from people. If this works, teams may spend less time coordinating and checking on AI helpers. Because the code is open, other developers can study it and build on it.

What to watch

Whether other developers try the prototype and report results.
How well it handles longer tasks without human help.
Whether the project grows beyond a research prototype.

Glossary

DeepSeek: A company that builds AI models and tools.
dsh-superagent: A research prototype that lets AI coding agents handle development requests on their own.
GitHub: A website where developers store code and work together on it.
Feishu: A team chat and work app, also known as Lark outside China.
MIT license: A open-source license that lets people use, change, and share the code with few limits.

Anthropic Manager Drops Coding Tests in Hiring as AI Outperforms Humans

Felix Rieseberg, Anthropic's recruiting manager, says the company has stopped using live coding tests in hiring because its Claude models now write code better than humans. Instead, he is looking at candidates' portfolios for taste, empathy and principled thinking, and values those who build prototypes with AI. Surveys cited in the article show AI is making technical assessment harder for employers, with system design, debugging and AI collaboration gaining weight.

AITechNodeNov 2

SenseTime open-sources 9-billion-parameter AI model that reuses layers for deeper processing

SenseTime has released SenseNova Loop-9B-Preview, an open-source multimodal model announced October 30 that reuses computation to extend processing depth without duplicating weights. It accepts text and images and generates text for document understanding, coding, and tool-use tasks, running a shared block of 24 decoder layers twice from a model of about nine billion parameters. SenseTime has published the weights, inference instructions, and training guidance under the Apache 2.0 license.

AIVnExpressNov 2

Speakers of Less Common Languages Pay More for AI

AI models consume more tokens when processing less common languages, putting those users at a disadvantage compared with English or other widely used language speakers. A CLEAR study cited in the article calls this the "hidden tokenizer tax," finding that among the 20 most common internet languages, the most expensive needed 1.5 to 5 times as many tokens as the cheapest, a gap that grew to a 17-fold difference across 200 languages.

AI전자신문Nov 2

Doubts Over AI Investment Returns Spread; Wall Street Hedges Risk with Put Options and CDS

Wall Street is moving to guard against losses from massive AI investment as doubts grow over whether returns will meet expectations. According to the Wall Street Journal, individual investors and tech executives are using put options and credit default swaps (CDS) to prepare for asset price declines, and demand for hedging against falling GPU rental fees is also rising. Put options on CoreWeave have surged, with trading concentrated in 2028-expiry, $35-strike contracts, while CDS premiums for five major tech companies rose over the past month to their highest levels since 2021.