DeepSeek open-sources a super-agent, ChatGPT and Codex merge usage limits, Karpathy ships HackerTube

dsh-superagent is a proactive coding agent that looks after a single GitHub repository. It runs on DeepSeek Harness (DSH), DeepSeek's open-source agent runtime. When an allowed person asks it for work in an issue or PR thread, it opens a work item, does the job in its own session and directory, and replies in the original conversation, for example with a pull request. It can also take requests from Feishu chats. Setup has many steps, so the README tells you to have your coding agent walk through SETUP.md with you.

git clone https://github.com/deepseek-ai/dsh-superagent-prototype

Then ask your coding agent: "Read SETUP.md in this repository and set up dsh-superagent with me."

If your team wants an agent that lives in a shared repo, this is a reference design for splitting permissions, execution and notifications. The agent acts only when an allowed person asks or a team-written rule allows it, and speaks only when someone needs to act, and docs 000 to 008 lay that out in order. The operating advice is worth borrowing as is: give the agent its own account, and start with proactive work only logged, not executed.

Starting in December, Plus, Pro and Business plans will draw ChatGPT and Codex from one usage allowance. The change comes as OpenAI folds the agentic features of Work into the main ChatGPT experience. Usage will be counted in tokens, the units a model reads and writes, so how fast you burn through it depends on the model and the task. Instant chat won't count, and the combined balance and reset times will show in Settings → Usage.

Today heavy Codex use leaves your ChatGPT allowance alone. From December, using one cuts into the other. Developers and teams who lean on Codex should measure now how much the heavier models and long tasks consume. Once you hit the limit, your only options are to wait, upgrade or buy credits.

HackerTube collects the YouTube videos that Hacker News commenters have kept recommending to each other for 20 years. Karpathy had Claude filter 43 million HN comments Vendor claim down to 6,454 videos Vendor claim, and says he built it in about two hours with Claude Code on Opus 5.5 at Medium effort. There's no login; state stays in your browser session, and thumbnails are a frame from the middle of the video. The comment analysis is downloadable as a single zip.

The point is the data: Karpathy invites you to point your own coding agent at the zip and build your own version. It's a compact example, from raw data to deployment, of using an LLM to filter a huge text corpus into a curation product, which maps well onto internal knowledge curation. Karpathy says his only manual steps were creating an API key, starting a VPS and buying the domain, a useful gauge of how much you can hand an agent.

Connected Apps Platform is a developer route for plugging outside services into the Gemini app and Google Search (AI Mode). The default path is a remote MCP server, the standard protocol agents use to call tools. On Android you can instead use Jetpack AppFunctions so the tools run on the device. Users link your app once and can then see their account data and take real actions from the conversation. Tool registration, usage analytics, review and publishing all happen in Google Connected Console, and the preview is taking applications.

Teams that already run an MCP server can reuse that integration to reach the Gemini app and Search, two large distribution channels. You choose whether your app shows up in Search, the Gemini app or both. Because it touches user accounts, plan for an OAuth 2.1 implementation up front.