Sun, Nov 1, 2026Sunday, November 1, 2026 · 11 stories · 4 min read

DeepSeek opens a coding agent that picks up work unprompted

Ramit KoulFounder, Software Engineer & Innovator · Published 5:02 AM ET

DeepSeek's Harness team has open-sourced a prototype coding agent that watches GitHub and team chats and starts on bugs without being tagged.1, 2 Other releases focused on the software around the model: Factory made programmatic tool calling the default in Droid 1.0, and Vercel Labs released agent-desktop, which lets agents operate desktop apps through accessibility trees.3, 4 The credentials those agents depend on are also being targeted, and Northflank told customers to assume all of their environment variables had been decrypted.5 Here is the story that matters most, then 10 briefs. Numbers in the text link to the references at the end.

Dev tools

DeepSeek releases an open-source Super-agent prototype that fixes bugs without being asked

Image: straitstimes.com

On October 31, DeepSeek's Harness team released a prototype called Super-agent, a more autonomous coding agent built to compete with Anthropic's Claude Code.1 It can watch developer group chats for reports of coding problems and suggest fixes before anyone calls on it.6 According to the GitHub repository, dsh-superagent-prototype is a proactive coding agent for a single GitHub repository, packaged as a bundle for DeepSeek Harness, and it can also listen to Feishu if you set that up.2

DeepSeek describes a two-layer architecture with event-driven workflows, asynchronous interactions across several channels, storage kept separate from compute, and an external watchdog.7 An internal instance has been running nonstop, reproducing and fixing bugs, responding to code reviews and helping develop DeepSeek Harness itself.7 DeepSeek says the goal is to take as little human attention as possible while giving as much value as possible.6

The README calls the code a research prototype, not an official DeepSeek product. It also says some designs in the accompanying write-up are not implemented yet.2 DeepSeek set up the Harness team earlier in 2026 to focus on coding agents, and it recently raised about 100 billion yuan ahead of a planned IPO.6

Why it matters for builders

The repository is a working reference for agents that react to chat and repository events on their own instead of waiting for a prompt.7 DeepSeek warns that the code may contain bugs and should be reviewed and adapted before anyone runs it in their own environment.2

In brief

  • Models

    Nace AI released open weights for Drex Max v1, its 27B decision model. Drex Max does not generate text: it scores every option for typed questions in a single forward pass, and a 4-bit build runs on Apple Silicon Macs with 24 GB of memory or more.8, 9 The weights use the Nace.AI Open RAIL-M license, and Nace reports its own score of 67.52 on Decision Index 0.4.8

  • Products

    A New York Times test found that ChatGPT for Teens still does homework and missed some warning signs. The reviewer found that the teen mode still gave direct answers and helped write essays, and that it missed some suicide warnings before responding properly later in the conversation.10 OpenAI launched the experience in August for users aged 13 to 17, and Common Sense Media has urged the company to pause the feature after its own tests.10, 11

  • Dev tools

    Factory shipped Droid 1.0 with programmatic tool calling turned on by default. Instead of calling tools one at a time, the model writes a JavaScript program that runs the tools in a sandbox, so a single request can drive hundreds of tool calls, and Factory says this cuts cost by 15% on Claude Opus 5.5.3, 12 In Factory's own Terminal-Bench 4 runs with Opus 5.5, Droid scored 65.2% against 57.6% for Claude Code, though with GPT-6.1 Sol it took 10% longer per passed task than Codex.3

  • Industry

    Northflank told customers to rotate every secret after stolen AI API keys were used. After a threat alert from Wiz on October 30, the hosting platform told customers to assume all environment variables had been decrypted, and it later said users had reported unauthorized use of AI keys stored on the platform.5, 13 Wiz counts at least nine incidents at AI developer platforms in October, including Railway, OpenCode and Warp, and says some attackers went after AI credentials to use in LLMJacking, which means running up AI usage on someone else's keys.5

  • Products

    Spotify is reportedly aiming to launch an AI remix and cover tool in early 2027. According to the Financial Times, Spotify may charge an extra $4 to $5 a month, has already signed licensing deals with Universal Music Group, Merlin and Believe, and is still negotiating with Sony Music and Warner Music.14 Artists would earn royalties when their songs are remixed and could veto use of their music.14

  • Infrastructure

    Darkbloom warned members that payouts from renting their Macs for inference will fall as token prices drop. Tom's Hardware, citing The Register, reports that the price per million tokens for Anthropic on the network fell 77%, from $2.65 in May to $0.62 in October, which would cut the median earnings of top operators to about $1.23 a day.15 Darkbloom says it will tell members which models are in demand and is looking at other workloads, such as distributed CI/CD jobs.15

  • Dev tools

    Vercel Labs open-sourced agent-desktop, a CLI that lets agents operate desktop apps through accessibility trees. The Rust tool gives an agent a JSON snapshot with a reference for every element it can act on, works on macOS, Windows and Linux, acts without moving the user's pointer by default, and shipped as version 1.0.2 under the Apache-2.0 license.4, 16 Its developer reports using 78% to 96% fewer tokens than full-tree dumps on dense Electron apps.4

  • Research

    Sam Altman said OpenAI will keep releasing AI-generated math results despite objections from mathematicians. On the Cheeky Pint podcast, Altman admitted OpenAI let things get out of control on math for a week, but he argued that humanity deserves fundamental knowledge.17 Critics say none of the hundreds of AI-generated proofs OpenAI released have been vetted, and the releases followed a disputed Navier-Stokes claim that cost around $10 million in compute.17

  • Research

    Experts told Engadget that Anthropic's warning about self-replicating small models is plausible but less new than it sounds. Anthropic's Frontier Red Team wrote that small models may be months away from being able to self-replicate on IoT devices such as off-brand Android TV boxes.18, 19 Georgia Tech professor Michael Bailey noted that the tests ran on an emulated network with a trivial exploit, and he expects AI to mostly change the economics between attackers and defenders rather than enable new kinds of attacks.19

  • Products

    Otter.ai is adding three AI features as it moves beyond transcription. The new features cover what CEO Sam Liang calls the whole meeting life cycle, and they come with $320 million in new funding from General Catalyst.20 The push deeper into the workday also raises familiar privacy questions.20

References

Every source behind this edition. Open one to read the full story.

  1. 1DeepSeek Drops Enhanced Coding Agent With Aim to Rival AnthropicBloomberg · bloomberg.com
  2. 2GitHub - deepseek-ai/dsh-superagent-prototype: Towards a More Autonomous Agent SystemGitHub · github.com
  3. 3Meet Factory Droid 1.0: A Model-Independent Coding Agent Harness Built on Script Modemarktechpost.com · marktechpost.com
  4. 4Vercel Labs Open-Sources agent-desktop for Agent Computer Usemarktechpost.com · marktechpost.com
  5. 5Northflank security breach leads to abuse of customers' AI API keysbleepingcomputer.com · bleepingcomputer.com
  6. 6DeepSeek drops enhanced coding agent with aim to rival Anthropicstraitstimes.com · straitstimes.com
  7. 7DeepSeek builds ‘super-agent’ for autonomous codingtechinasia.com · techinasia.com
  8. 8Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Siliconmarktechpost.com · marktechpost.com
  9. 9nace-ai/drex-max-v1 · Hugging FaceHugging Face · huggingface.co
  10. 10Why my conversations with ‘ChatGPT for Teens’ made me very worriedstraitstimes.com · straitstimes.com
  11. 11ChatGPT for Teens Risk Assessmentinstitute.commonsensemedia.org · institute.commonsensemedia.org
  12. 12Droid 1.0: Building a Frontier Harness | Factoryfactory.com · factory.com
  13. 13Security incident - Incident details - Northflank - Statusstatus.northflank.com · status.northflank.com
  14. 14Spotify Eyes Early 2027 Release for AI Remix Toolpcmag.com · pcmag.com
  15. 15Firm that lets you rent out your idle Mac for AI inference workloads warns of plummeting token costsTom's Hardware · tomshardware.com
  16. 16GitHub - vercel-labs/agent-desktop: Agent Desktop gives any agent reliable computer use on the desktop. Built with Rust, it sees any app's real UI structure through OS accessibility trees and operates it — refs stay stable and actions stay safe to retry, instead of guessing from pixels.GitHub · github.com
  17. 17Sam Altman to Infuriated Mathematicians: The Math Slop Won’t StopFuturism · futurism.com
  18. 18AI models may be capable of self-replicating without access to a datacenterAnthropic · anthropic.com
  19. 19What happens when AI starts self-replicating?Engadget · engadget.com
  20. 20Otter.ai is moving beyond transcription—and deeper into your workdayFast Company · fastcompany.com

That’s the edition.

The next one is ready tomorrow by 6 AM ET.

Written with AI from the sources in the references above.

Ramit Koul
AuthorRamit KoulFounder, Software Engineer & Innovator