DeepSeek Open-Sourced an Always-On Coding Agent. Its Workers Can Run Commands Without Approval, and It Was Only Tested on Simulated Tools.
deepseek-ai/dsh-superagent-prototype is an MIT-licensed TypeScript bundle for DeepSeek Harness: a Router plus isolated Worker sessions bound to one GitHub repository, with optional Feishu, that turns allowed requests into work items and pull requests. The README calls it a research prototype for reference only, not an official DeepSeek product, and says much of it was written by coding agents. In auto accept mode, Workers can run commands as the operator's OS user without per-command approval. The latency figures circulating, a p95 of 15.7 minutes against 29.5, come from MarkTechPost on synthetic messages and simulated tools. We read the repository.
By T.W. Ghost
DeepSeek has published an always-on coding agent, and the first line of its README is a disclaimer. deepseek-ai/dsh-superagent-prototype is an MIT-licensed TypeScript bundle for DeepSeek Harness: a Router that takes requests, and isolated Worker sessions that do the work, scoped to one GitHub repository with optional Feishu as a second channel. Allowed requests become work items and then pull requests. The banner at the top: "Research prototype for reference only, not an official DeepSeek product." The README also says much of the engineering was done by coding agents, and warns of bugs.
Two things decide whether you should run it. The first is the accept mode: off, observe, which is the default, or auto. In the auto paths, Workers can run commands as the operator's OS user without a per-command approval. The second is what the only performance numbers are made of. MarkTechPost reports a p95 of 15.7 minutes against 29.5 for a single-session baseline; those runs used synthetic messages and simulated tools, and they are not on any DeepSeek-owned page we could find. They measure the router's scheduling, not the code it produces. We read the repository README. Here is what held up.
What Is in the Repository
| dsh-superagent-prototype | |
|---|---|
| Publisher | deepseek-ai on GitHub |
| License | MIT |
| Status, per README | Research prototype for reference only; not an official DeepSeek product; much of it written by coding agents |
| Design | Router plus isolated Worker sessions; one GitHub repo; optional Feishu |
| Output | Work items and pull requests |
| Accept modes | off; observe (default); auto |
| Worker rights | In auto paths, commands run as the operator's OS user without per-command approval |
| Latency claims | p95 15.7 vs 29.5 minutes (MarkTechPost; synthetic messages, simulated tools) |
| Related billing | DeepSeek Search and Fetch at $0.75 per 1,000 requests, on DeepSeek's pricing docs, if you wire the agent to live tools |
"Not an official product" is the vendor's framing, and it is the right one. This is reference code for a harness, released so people can read it. Treat it as a design document that happens to run.
Observe is the default for a reason. The step from observe to auto is the step from an agent that proposes to an agent that executes with your shell rights. Read the setup and authorization docs before you take it.
The minutes are not a quality claim. A faster p95 on simulated tools says the router hands out work efficiently. It says nothing about whether the pull requests are good, because nobody has measured that.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Public repo under deepseek-ai; MIT license | Verified (GitHub) |
| Research prototype, not an official product; agent-written; bug warning | Verified (README) |
| Router and Worker design; one GitHub repo plus optional Feishu | Verified (README) |
| Workers can run commands without per-command approval in auto mode | Verified (README accept-mode description) |
| p95 15.7 vs 29.5 minutes | Press (MarkTechPost); not on DeepSeek docs; simulated tools |
| Search and Fetch $0.75 per 1,000 requests | Verified on DeepSeek pricing docs (related, not the agent repo) |
Who Should Use It
Read it if you are building an always-on agent and want to see how a lab structures a Router and Worker split; the code is the value.
Skip running it in auto mode on a machine whose OS user has anything you care about, and skip quoting the MarkTechPost minutes as a DeepSeek SLA.
The frontier is elsewhere: the always-on agents that are products this fall, Grok Bot and Dots, ship with the same question this prototype makes explicit, which is how much of your machine the worker gets. Our model comparison has the current standings on the models underneath.
Sources
- GitHub: deepseek-ai/dsh-superagent-prototype
- MarkTechPost: DeepSeek dsh-superagent Router and Worker coding agent (press, for the latency figures)
- DeepSeek API docs: pricing
*Not sure how much shell access an always-on agent should have on your repo? Take the free 2-minute quiz and get matched. Then read about the model this harness is built around and the swap that moved its score: DeepSeek V4.1 Flash 1017.*