DeepSeek's dsh-superagent: One Ticket Per Issue, Multiple Agents Run GitHub Workflows in Parallel

DeepSeek has put a research prototype on GitHub, and the repository is called dsh-superagent-prototype. After Leandro von Werra reposted it on X, Eric Hartford also reposted it. The project description is direct: multiple agents process GitHub issues and PRs in a repository in parallel, with the goal of minimizing human attention while producing as much value as possible. The documentation subtitle is even more specific: One Router Session, One Worker Per Ticket, As Little Human Attention As Possible. It is explicitly labeled as a research prototype, not a product.

How It Runs

GitHub and optionally Feishu turn events into triggers, which first pass through a deterministic gate that checks whether the sender is on an allowlist. The router in the main session evaluates the remaining content and records a decision: accept, stop, ignore, or unavailable. Acceptance creates a ticket with immutable instructions. After that, each ticket runs in its own work session and directory, usually a git worktree.

Figure 1 in the documentation shows the layout: one router, isolated workers, replying on the same thread. Figure 2 shows the ticket lifecycle, time limits, and acceptance modes. This structure breaks the issue and PR workflow into an auditable chain of events, with code generation being just one part of it.

Details Matter More Than Slogans

The Data section lists a set of limiting parameters: number of work slots, evaluation timeout, ticket preparation command, time limit after background work, and so on. They show that the project is controlling concurrency and timeouts to prevent agents from expanding without bound. At the bottom of the documentation there is also a security warning: operators should run the agent under a dedicated operating system account and protect credentials.

From an engineering perspective, how smart the agent is can be set aside for now; how permissions, isolation, and timeouts are implemented is more worth examining. Human attention is treated as a scarce resource: it is first filtered by the gate and router, then handed to parallel workers to try. For development teams, the value may lie in automating low-value triage and parallel attempts while leaving an auditable record of decisions. The risks are also clear: an allowlist, credential isolation, concurrency slots, and timeouts—if any one of them is loose, it can turn into an incident.

Project address: https://github.com/deepseek-ai/dsh-superagent-prototype

Source: Leandro von Werra's X post and the dsh-superagent project documentation.