XNews

DeepSeek Open-Sources dsh-superagent Autonomous Coding Agent Prototype

DeepSeek Open-Sources dsh-superagent Autonomous Coding Agent Prototype

DeepSeek open-sourced dsh-superagent, a proactive coding agent prototype for a single GitHub repository packaged as a bundle for DeepSeek Harness. The agent converts incoming requests into persistent tasks with isolated sessions, working directories, branch state, execution history and pending inputs, gated by an authorization layer before any work begins. It separates reactive and proactive behavior, pausing tasks for clarification and recovering ongoing work after failures or restarts.

Event version 1.

Key insights

Persistent task model

Incoming work is turned into persistent tasks rather than separate coding sessions. Each task gets its own agent session, working directory, branch state, execution history and pending inputs, allowing multiple tasks to run independently without mixing context.

Authorization before execution

An authorization layer requires a request to come from an allowed source or match an explicit rule before the agent can act. Once authorized, the task is assigned to an executor that can inspect the repository, make changes, and return results through the original conversation or open a PR.

Pause and recovery

If the agent needs clarification, it can pause while keeping the same task and session alive, and follow-up messages route back to that existing context. The system retains enough task state to recover ongoing work after failures or service restarts.

Reactive versus proactive behavior

The agent does not continuously invent work for itself. It acts when explicitly asked or when a written rule grants permission to act proactively.

Autonomy from system design

Autonomy comes from persistent task state, isolated agent sessions, authorization rules, concurrent execution, recovery and GitHub-native outputs working together around the model, not just from giving a coding model access to tools.

Source posts