DeepSeek Harness Grew Two Ends: One Waits to Be Asked, One Fits in a Pocket

Two repositories in the DeepSeek Harness ecosystem crossed twenty thousand stars within a week of each other, and they solve opposite problems.
deepseek-ai/dsh-superagent-prototype is a proactive coding agent that watches a single GitHub repository and does work without being watched. ccch1mneyyy/dsh-Go is an Android client that puts the entire harness on your phone, or lets the phone drive the copy on your desktop.
One is built to take almost none of your attention. The other makes the harness impossible to leave behind.
The superagent's real design constraint
The README opens with a disclaimer before it opens with a feature. This is a research prototype, not an official DeepSeek product. It grew out of the team's day-to-day development work, much of the engineering done by coding agents, and it may contain bugs. Review it before pointing it at your own environment.
Then it states the philosophy in one line: take as little human attention as possible, and give as much value as possible.
The follow-up sentence is the one that matters. An agent that can act for a whole team can also waste a whole team's attention, so it acts only when asked or when a rule the team wrote allows it, and speaks only when someone must act or can use a result.
That's a design position, not a marketing line, and it shapes the whole system. The agent listens to GitHub, and optionally Feishu. When someone it's allowed to take instructions from asks for work, it opens a work item, spins up its own agent session and working directory, does the job, and replies in the original conversation. A pull request is the typical answer.
A web panel exists for one purpose: showing what needs a person. Not activity, not logs, not a chat window. The queue of decisions nobody has made yet.
Setup is delegated to an agent
The installation instructions are a sentence you paste into a coding agent: read SETUP.md in this repository and set up dsh-superagent with me.
That's a deliberate choice, and a slightly unnerving one. Setup involves enough steps that a human following a list would get lost somewhere around OAuth scopes. The agent asks one short question at a time, shows the full plan once, then installs and configures everything.
The documentation runs 000 through 008 in a numbered sequence the README asks you to read in order. Philosophy, domain model, authorisation, execution, prompts, panel, deployment, architecture, cookbook. The authorisation document is third, ahead of execution and prompts, which tells you where the team thinks the risk lives.
pnpm check runs the typecheck, tests and build. MIT licence.
The phone client does the opposite
dsh-Go hit number one on Trendshift's all-language daily chart on 30 October and has 22,418 stars. It is an Android client, AGPL-3.0, Android 8 and up.
The pitch is in the tagline: run it on the phone itself, or use the phone to drive the DSH on your computer. One tap switches between the two.
It also ships a built-in Linux environment with no root and no Termux. That's the part that makes local execution plausible rather than a demo. A harness needs a shell, a filesystem and the ability to run processes; getting those on Android without root is normally the reason projects like this die at the prototype stage.
Installation is an APK from GitHub Releases or the project's own download page, with arm64 as the default for most phones. The author says iOS and HarmonyOS adaptations are in progress and jokes that more stars will speed them up. Uninstalling wipes conversations, the Linux environment and settings, and a differently-signed early build can't be installed over a newer one, which is the sort of warning that only gets written after someone loses a week of work.
There are also two animated mascots, Deepy and a whale-girl, that type while the agent works, do a flip when a job finishes, and raise a hand when they need confirmation. There's a toggle to make them follow you onto the desktop. It's easy to dismiss as decoration, and it's also the clearest signal that this project thinks of the agent as something present rather than something queried.
What the pair says about the harness layer
Put the two repos side by side and the shared assumption is obvious. Both treat the harness, the prompts, tools, memory and permission rules around the model, as the product. Neither is interesting because of the model underneath.
The split is about where attention goes. The superagent's whole design is a negotiation over when it's allowed to interrupt, and the README points at a document called philosophy to explain what that asks of the operator. The phone client's design is the opposite bet: if the agent is always with you, you'll supervise it more, so make supervision cheap.
Both positions have a cost that the READMEs are honest about. The superagent only makes sense for one repository, which is a narrow scope and an intentional one. The phone client runs a Linux environment on a device with a battery, and nobody has published what a long agent session does to it.
There is a governance asymmetry too. The superagent's authorisation model assumes a team that can agree on rules and change them when someone leaves. The phone client assumes one person holding one device, where the permission boundary is whatever that person is willing to tap. Two ends of the same harness, two completely different answers to the question of who is allowed to say yes.
Who should try which
If you maintain a repository with steady, well-scoped work and a team that already writes rules down, the superagent prototype is the more useful of the two. It's the closest thing to an agent that owns a queue rather than answering prompts.
One prerequisite deserves a line of its own. The whole model depends on someone having written down what the agent may do unsupervised. A team that currently approves every change in conversation has no rules to hand it, and a team that has rules probably wrote them for people. Translating those into conditions a machine can evaluate is the actual work of adopting this thing, and the prototype won't do it for you.
If you want to watch an agent work while you're away from a desk, dsh-Go is the more immediately satisfying. The remote-control mode is the underrated half: pointing your phone at the desktop harness gets you the phone experience without giving up the desktop's CPU.
Neither is production software. Both reward an hour of reading. The superagent's documentation earns it even if you never install anything, because the argument about attention is the argument every team deploying an agent will have in the next year.
Related articles
Google replaced the model behind Search, and the number that moved was a shell benchmark
Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.
Three ways a coding agent wastes money, measured on 30 repositories
TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.
earshot gives coding agents 922 sounds, and every one is rendered from code
An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.
The agents stealing API keys were not sophisticated. The gateways were open.
Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.