DeepSeek releases a super-agent prototype for parallel team coding
Edited and reviewed by AI Ping Editorial Team
3-line summary
- The research prototype connects one GitHub repository and optionally Feishu conversations.
- Authorized tasks run in separate sessions, with results and decisions returned to the original thread.
- Inferred work defaults to observation; execution sessions use the operator’s machine permissions.
DeepSeek's Harness team has published dsh-superagent, a research coding prototype that accepts team requests and runs work in separate agent sessions. It watches one GitHub repository and can optionally connect to Feishu, the workplace messaging service. Authorized requests can produce a proposed code change, with a reply sent back to the conversation where the work began.
GitHub's public metadata dates the repository's creation to October 31. The team labels the release research code rather than an official DeepSeek product. It adapted an internal implementation for publication, and says some ideas in its accompanying design article describe future work. The public code therefore has a narrower scope than the complete design discussion.
The organizing unit is a piece of team work. One session assesses incoming requests, while separate sessions carry them out. Each accepted task records its permitted scope and the conversations that should receive results. A conversation can contain multiple tasks, and a task can follow other discussions for context.
Reading a conversation does not authorize new work
An allowed person who directly addresses the agent creates an explicit request. Starting work inferred from other messages requires the team's written rules and the relevant acceptance mode. The default observation mode records what the agent would have accepted, without turning that judgment into a new task.
Messages from related conversations supply context without granting authority. Workers handle follow-up requests within their scope and hand back requests that exceed it for another routing decision.
The team's stated design priority is the human attention spent supervising agents. Its communication rules call for one acknowledgement, quiet work in progress, and another message when a result is useful, a decision is needed or the job cannot finish. Replies go to the original conversation, preserving where colleagues expect to find the outcome.
Each task receives its own session and directory. A concurrency limit controls worker slots, with an extra slot for a human reply or status request. This reserves a way to continue existing work even when new jobs fill the queue.
Teams can separately edit communication, task scope and delivery requirements, including evidence attached to pull requests. Default instructions require drafts to remain identified as unfinished: opening a pull request alone does not establish readiness. Interface changes require a real before screenshot. Authorization gates and tool checks remain enforced in code, separately from editable prompts.
The supervision panel prioritizes unanswered questions, results awaiting review and blocked or failed work. Individual records expose the originating request, authorization scope and execution assignments. The panel also supports retrying failed work and pausing processing. Its documented interface principle is to give healthy states no space, concentrating attention on items that need intervention.
Decisions retain reasons and matching rules. Filtered or dropped messages link to their conversation or task, while search accepts comment links and issue references. These records let supervisors trace how a request was handled.
Execution sessions run commands as the operator’s operating-system user without approval. The execution document explicitly says directories and command wrappers establish defaults, not a security boundary. Machine access to files and credentials therefore requires separate consideration from the policy that authorizes a coding task.
The MIT-licensed code requires a configured model provider and a dedicated agent GitHub account with repository write access. That account posts comments, pushes branches and opens pull requests. Deployment documentation supports repositories on github.com, excluding GitHub Enterprise Server. Optional Feishu integration requires a bot. Operators configure the repository, accounts, allowed people and execution locations through the panel; setup must pass before polling or work starts.
Sources
- dsh-superagent research prototype repository — DeepSeek Harness Team
- Team tasks, conversations and authorization scope — DeepSeek Harness Team
- Explicit requests, inferred work and observation mode — DeepSeek Harness Team
- Human attention and communication rules — DeepSeek Harness Team
- Worker sessions, concurrency and execution permissions — DeepSeek Harness Team
- Supervision panel and task records — DeepSeek Harness Team
- Team instructions and reviewable delivery requirements — DeepSeek Harness Team
- Deployment prerequisites, agent accounts and supported repositories — DeepSeek Harness Team
- MIT License — DeepSeek Harness Team
- Public repository metadata — creation date — GitHub