Your repository
The shared project, its issues, pull requests, reviews, and CI context. The prototype is designed around one repository per deployment.
A research prototype by deepseek-aiDeepSeek SuperAgent:
DeepSeek SuperAgent:
less attention, more progress.
Meet DeepSeek SuperAgent. A proactive coding agent that turns your team’s requests into work, keeps the context, and comes back when it has something that matters.
Needs your attention
1Useful results. A clear next step.
Awaiting reviewDEMO-01
Fix the flaky authentication test
Fix ready for your review
agent/fix-auth-testPanel guideFROM REQUEST TO RESULT
@agent can you investigate this failing test?GitHub · Allowed teammate
Authorization & scope checkedEval → Ticket
Investigating the failing testDedicated session & directory
Your rules. Your repository. Your oversight.
Quiet until it matters.Designed around human attention
ONE CONNECTED WORKFLOW
GitHub
+ Feishu optional
DeepSeek Harness
You, in control
WHAT IS DEEPSEEK SUPERAGENT?A coding agent that works
A coding agent that works
with your whole team.
DeepSeek SuperAgent connects team conversations with coding work in one GitHub repository. Built on DeepSeek Harness, it receives requests, assigns tasks, and returns results for review.
Mention DeepSeek SuperAgent in an issue, pull-request discussion, or configured Feishu chat. It checks the requester and permitted scope, then opens a ticket with its own session and working directory.
DeepSeek SuperAgent is designed to ask for attention when there is something to act on: a result to review, a concrete question, or a blocker. The web panel collects those items, while the original conversation remains the place to discuss the work.
DeepSeek Harness
The underlying agent runtime that provides sessions, models, tools, and the execution environment used by the workers.
SuperAgent
The coordination layer: receive requests, evaluate authorization, route tickets, deliver results, and make human oversight practical.
DeepSeek SuperAgent is published by deepseek-ai as a research prototype, not an official DeepSeek product. The public implementation also differs from the team’s internal version and research article. README
DeepSeek SuperAgent featuresA teammate, with a
A teammate, with a
well-defined role.
More than a chat window. A structured way to receive work, act within boundaries, and deliver something useful.
01
Meet your team where it works
DeepSeek SuperAgent takes requests from GitHub issues, pull requests, and comments. Add Feishu when your team works in chat.
GitHub + optional Feishu02
Autonomy, with a clear scope
Allowed lists and team policies determine who can request work and what it may include. Each accepted task records that scope.
Explicit authorization03
A workspace for every task
Each ticket gets its own session and working directory. Configure how many workers can run and how each checkout is prepared.
Dedicated sessions04
Close the loop, in context
Results return to the original conversation, for example as a pull request. Follow-up comments reach the same task through its subscribed threads.
From request to delivery05
Human attention comes first
The DeepSeek SuperAgent panel puts questions, blocked work, and review-ready results first. Pause processing and inspect task details.
Meaningful oversight06
A workflow you can inspect
Follow a request from its first message to evaluation and execution. Review decision reasons, policy versions, and prompt history.
Traceable decisionsDeepSeek SuperAgent use casesFrom a team conversation
From a team conversation
to a reviewable change.
These DeepSeek SuperAgent workflows are illustrative examples, not benchmark results. What the agent can complete depends on your repository, model, tools, and configuration.
Issue → implementation
Investigate a bug without losing the thread
An allowed teammate asks DeepSeek SuperAgent to investigate a reproducible issue. The executor reads the repository, works in its ticket directory, and returns a result to that conversation. If the investigation needs a decision or cannot continue, the ticket makes that human action visible.
@agent investigate the failing login testReview → follow-up
Keep a pull request moving
Review feedback on a DeepSeek SuperAgent pull request reaches the worker through its subscribed thread. The worker handles requests within the ticket’s scope and returns unrelated requests for separate evaluation.
@agent address the review commentsFeishu → repository
Bring chat requests into the workflow
Mention the DeepSeek SuperAgent bot in a configured Feishu chat to request repository work. The same authorization checks apply, and results or questions return to that conversation.
@agent add an activity summary“
THE PRINCIPLE BEHIND THE PROTOTYPE
Take as little human attention as possible, and give as much value as possible.DeepSeek SuperAgent philosophy
How DeepSeek SuperAgent worksThe work moves forward.
DeepSeek SuperAgent architectureThe work moves forward.
You keep the overview.
01
A request comes in
Mention DeepSeek SuperAgent on GitHub or in a configured Feishu conversation. A source records the trigger.
Source → Trigger02
The scope is checked
The evaluator checks the requester and team policy. Accepted work becomes a ticket with a defined scope.
Eval → Ticket03
The agent gets to work
A worker acts in its own session and directory, receiving follow-up context as it arrives.
Assignment → Executor04
A useful result returns
DeepSeek SuperAgent sends results back to the conversation. The panel highlights review requests, questions, and blocked work.
Channel → Your teamDeepSeek SuperAgent authorizationChoose how much
Choose how much
initiative to give it.
DeepSeek SuperAgent handles explicit requests and inferred opportunities through different paths. The acceptance mode controls inferred work; explicit authorized requests work in every mode.
A mandate for every accepted task
DeepSeek SuperAgent uses an allowed list to identify who may request work. Source settings choose which conversations it receives. Each incoming event is then treated as a direct request, a possible task to assess, or something to ignore.
An accepted ticket records its authorized scope, called a mandate, and the policy version used. Subscribed threads add context without extending that scope. Before completed work resumes, the host checks that its authorization still holds.
Observe first. Automate deliberately.
DeepSeek SuperAgent starts in observe mode. Review its proposed tasks, reasoning, and matched rules before enabling automatic acceptance.
In DeepSeek SuperAgent auto mode, a proposed task needs a positive assessment and a matching team rule. Messages and quoted documents cannot change those permissions.
offExplicit work only
Inferred messages do not become candidates. Use this when you want the agent to work from authorized requests without proactively assessing other activity.
observeLearn before enablingDEFAULT
The default. Possible tasks are assessed and positive decisions are recorded as “would accept,” without starting the work.
autoAct within team rules
Possible tasks can start after a positive assessment and a match with at least one enabled rule. Explicit requests still work, and decisions are recorded.
DeepSeek SuperAgent supervisionWhat needs me now?
What needs me now?
Start with that.
The DeepSeek SuperAgent panel lives in the DeepSeek Harness web sidebar. It prioritizes work needing a person, then system health, then the history behind a piece of work.
Questions & review
See DeepSeek SuperAgent tickets awaiting an answer, ready for review, or blocked on human action.
Health & activity
Check setup issues, recent activity, connected sources, and how many workers are busy.
Controls & history
Pause processing, retry eligible work, inspect decisions, and trace a task back to the message that started it.
| Panel tab | What you can inspect |
|---|---|
| Overview | Tickets needing attention, system problems, the last 24 hours of activity, and executor load. The summary links to the relevant detail pages. |
| Source | The state of GitHub and Feishu inputs, setup problems, filtered triggers, and dropped messages. |
| Trigger | Conversation threads, incoming triggers, decisions, and subscriptions. Follow the events that led to a piece of work. |
| Eval | Recorded decisions with their reasons and matched rules. Inspect why work was accepted or ignored. |
| Execution | Executor state and ticket stages, waiting inputs, assignments, and actions such as pause, retry, or stop where available. |
| Settings | Source configuration, evaluation policy, execution settings, team prompts, and deployment details, with save history. |
Configure DeepSeek SuperAgentYour standards.
Your standards.
A persistent workflow.
DeepSeek SuperAgent supports continuity through team instructions, clear handoffs, and recovery when work is interrupted.
Teach it how your team works
DeepSeek SuperAgent team prompts define Communication, Ticket scope, and Delivery requirements: when to speak, what work is allowed, and what evidence to deliver. Team skills provide additional knowledge when needed.
Prompt changes are versioned. Sessions read the latest configuration with each model request, so edits apply without restarting a session. The host still enforces authorization separately.
Team prompts & skillsHandle interruptions explicitly
A blocked or failed DeepSeek SuperAgent ticket explains what needs attention. A human reply or eligible retry can continue the work. After a restart, replacement assignments receive recovery notes.
A person’s pause remains in effect until a person lifts it. Directory separation and credential wrappers establish working defaults, not an OS sandbox; deployment still needs an environment you are prepared to let the agent act in.
Execution & recoveryGet started with DeepSeek SuperAgentStart with the source.
Start with the source.
Set up with an agent.
Set up DeepSeek SuperAgent with help from your coding agent. Clone the prototype, open the project, and follow SETUP.md to choose the repository, accounts, and permissions.
DeepSeek SuperAgent setup guideBring a compatible DeepSeek Harness runtime, a model provider, and a dedicated GitHub agent account. The guide lists all prerequisites.
YOUR FIRST CONVERSATIONSETUP.md
1
Clone the research prototype
git clone https://github.com/deepseek-ai/dsh-superagent-prototype.git
cd dsh-superagent-prototype2
Ask your coding agent
Read SETUP.md in this repository and set up dsh-superagent with me.3
Review the plan, then configure
Choose your repository, sources, and allowed teammates. Confirm the setup plan before installation.
Research code, not production-ready software. Review and adapt before use. Workers execute commands as your OS user.
DeepSeek SuperAgent documentation
Read. Understand. Build.
Read DeepSeek SuperAgent’s primary sources to understand the implemented features and the prototype’s limits.
DeepSeek SuperAgent FAQGood questions.
Good questions.
Clear answers.
Understand what DeepSeek SuperAgent does, where it fits, and what to know before you run it.
What is DeepSeek SuperAgent?
DeepSeek SuperAgent is a research prototype built on DeepSeek Harness. It handles coding tasks for one GitHub repository, with optional Feishu integration.
Is this an official DeepSeek product or website?
No. The deepseek-ai repository identifies the code as a research prototype, not an official product. This independent guide is not affiliated with or endorsed by DeepSeek.
How is DeepSeek SuperAgent different from DeepSeek Harness?
DeepSeek Harness provides the runtime. DeepSeek SuperAgent adds request handling, authorization, task coordination, result delivery, and a supervision panel.
Does DeepSeek SuperAgent act on every message it sees?
No. Allowed lists and team policy determine which messages can start work. Subscribed conversations add context without granting more authority; proactive tasks also need matching rules.
Is it production-ready or sandboxed?
It is research code that needs review and adaptation. Workers run commands as the operator’s OS user without per-command approval. Separate directories are not a security sandbox; review the deployment guide before use.
Is it free to use?
The code is MIT-licensed. Model API usage, hosting, and third-party services may incur their own charges. A separately operated hosted trial is planned; join the waitlist for its launch notification. The trial is not available yet.
Can I deploy DeepSeek SuperAgent on Vercel?
This website runs on Vercel. DeepSeek SuperAgent needs a separate, long-running Harness instance with local state and worker processes. Follow the official SETUP.md and deployment guide.
Give your team room to focus.
Discover how DeepSeek SuperAgent connects repository work with team conversations.
Coming soon · Join the waitlist