DeepSeek dsh-superagent: a coding agent that watches team chats
Key takeaways
- dsh-superagent is an MIT-licensed research prototype that handles requests for one GitHub repository and can also listen to Feishu
- DeepSeek's simulation shows a p95 task time of 15.7 minutes against 29.5 minutes for a single session, but only on synthetic messages and simulated tools
- It runs with full OS-user access, so it needs isolating before anyone tries it
- Proactive work defaults to observe mode, where the agent only logs what it would have accepted
DeepSeek has released dsh-superagent, an experimental coding agent that takes requests from a software team's GitHub and, optionally, Feishu (ByteDance's workplace chat app) and does the work in separate sessions. It is MIT-licensed, and aimed at Anthropic's Claude Code, according to Bloomberg's reporting. DeepSeek itself calls it a research prototype and "not an official DeepSeek product".
What it does well
The design goal is stated in one line in the project's manual: "Take as little human attention as possible, and give as much value as possible." Most coding agents treat a single session as the basic unit. DeepSeek's team argues that a person can actively follow only 3 to 5 sessions at once, so attention, not model capability, becomes the bottleneck when running several agents.
The answer is a Router-Worker split. A long-running Router handles interaction and dispatch. Isolated Worker sessions do the actual work. In practice, an approved person asks for something in GitHub or Feishu. The agent opens a Ticket, which is a unit of authorised work carrying an immutable Mandate that records why it may proceed. Each Ticket gets its own session and directory. Results come back in the original thread, often as a pull request.
Several details suit real team use. A Subscription feature means later comments on a pull request reach the right Worker without a fresh mention, so reviewers do not have to re-summon the agent. The agent confirms once, stays quiet while working, and writes again only with a result, a question or a failure. A local web panel lists only the items that need a person. Shell commands move to the background after 5 seconds, so Workers keep receiving messages while a build or test run grinds on.
The restraint is deliberate. DeepSeek warns that an agent acting for a whole team can waste the whole team's attention, so it acts only when asked or when a team's own rule allows it. Proactive behaviour is gated by three accept modes: off, observe (the default) and auto. In observe, the agent merely records what it would have accepted, which lets a team audit its judgement before granting it any autonomy.
Where it falls short
The performance evidence is thin. DeepSeek tested with simulated group chat and GitHub CI streams, each run containing about 300 messages, of which about 20 were real work requests. Each scheme ran 5 times. The agent's p95 task completion (the time within which almost all tasks finished) was 15.7 minutes, against 29.5 minutes for a single-session baseline, about 1.9 times faster. Its p50 was 12.4 minutes; the baseline had no p50 reported.
Read those numbers carefully. They are self-reported, drawn from secondary accounts of DeepSeek's own write-up, and tested only on synthetic messages and simulated tools. Real coding quality was not measured. The baseline also looks weak: under the neutral prompt it called a subagent zero times in 5 runs, so the comparison partly shows that a single session which never delegates is slow.
There is also the wider problem of agents working together. In January, researchers behind the CooperBench study found that coding agents did worse as a team, with success rates on average 30% lower than when each agent worked alone. That study did not test DeepSeek's agent, but it frames the open question: dispatching in parallel is not the same as cooperating well.
Security is the other caveat. The agent runs with full OS-user access, so it needs isolating before use. The public code also differs from DeepSeek's internal version, and some ideas in the design post are not built yet, including an external Watchdog, separation of storage and compute, and an asynchronous ReAct loop. Anyone weighing it should treat the setup much as they would a checklist for giving an AI agent real access: least privilege first, trust later.
How it compares
| Tool | Model | Approach | Status |
|---|---|---|---|
| DeepSeek dsh-superagent | Whatever the Harness profile is set to | Router dispatches Tickets to isolated Workers; one repository; GitHub and optional Feishu | v0.1.0 research prototype, MIT licence |
| Claude Code (Anthropic) | Not stated in the source | The product DeepSeek is aiming at, according to Bloomberg | Not detailed in the source |
| GitHub Copilot app | Not stated in the source | Runs agent work across several repositories in parallel sessions; shown at GitHub Universe | Shown at GitHub Universe on Wednesday |
| DeepSeek agent app | Not stated in the source | Handles coding, documents and spreadsheets; preview released in August | Open source, public preview |
The distinction is scope. GitHub's Copilot app spans multiple repositories; dsh-superagent is deliberately limited to one. Its pitch is less about raw coding ability and more about how a team interacts with several agents without babysitting them.
Pricing and value
The code is free under the MIT licence, which permits commercial use and modification. Costs come from elsewhere: the model behind it, which is whatever the Harness profile is set to, plus the machine it runs on. The prerequisites are modest but specific: Node.js 22.19+ or 24+, git 2.32+, Python 3.8+, tmux and GitHub CLI 2.48+.
One cost shows up in the ablation figures. Removing direct event pass-through raised Router decisions by about 30% and cache-read tokens by about 37%, from 23.23M to 31.77M.
UK and US considerations
For US readers, the release lands as civilian federal agencies there were told to remove DeepSeek last month, per Bloomberg. Public-sector and contractor teams should check their obligations before even evaluating it. The company has also raised about $15bn ahead of a planned listing, Bloomberg reported.
The source reports no ICO action.
Who should use it suits engineering teams already working in GitHub who want to experiment with ticket-style delegation to agents, and who have a sandbox machine to spare. Platform and DevOps teams interested in agent orchestration patterns will find the Router-Worker design, the Mandate concept and the observe mode worth studying even if they never deploy it.
It does not suit anyone wanting a production tool. It is v0.1.0, released as one commit, with unproven real-world coding quality.
The useful ideas here are the authorisation model and the attention-saving behaviour, not the benchmark. Leaving the agent in observe mode first, reviewing what it would have accepted, then widening its remit is a sensible template for any proactive agent.
Verdict
It is not worth deploying against live repositories yet. The 1.9x speed claim rests on simulation and a weak baseline, and nobody has measured whether the pull requests it produces are any good. Watch for independent testing on real repositories, and for which of the unbuilt ideas, particularly the Watchdog, DeepSeek ships next.
Frequently asked questions
What is dsh-superagent?
An MIT-licensed research prototype from DeepSeek that works on one GitHub repository. It takes requests from approved people, runs each in its own session and returns results, often as a pull request.
Is dsh-superagent safe to run?
DeepSeek calls it a research prototype and tells users to review and adapt it first. It runs with full OS-user access, so it should be isolated before use.