DeepSeek Super-Agent: How the Open-Source Coding Prototype Works

DeepSeek's Harness team published Super-agent on October 31, an open-source research prototype that turns GitHub requests and optional Feishu messages into coding work. It can keep track of several tasks and return results to their original conversations. Developers can inspect the code, but DeepSeek warns it is not production-ready.

The useful change is continuity. A bug report, a code review comment and the next test result can become inputs to the same piece of work, without a person repeatedly copying them into a chat.

What DeepSeek released

The dsh-superagent-prototype repository is MIT-licensed code packaged for DeepSeek Harness. It works on one GitHub repository, with Feishu as an optional source of requests. Accepted work gets its own agent session and working directory, and a web panel shows where a person is needed.

This is a research prototype, not an official DeepSeek product. The team says the public version differs from its internal implementation, and some ideas in its accompanying design discussion are not implemented. That distinction matters when reading descriptions of an agent that independently monitors a whole team's work.

There is no new model or dedicated Super-agent API price attached to this release. Running it requires a configured model provider and an environment for DeepSeek Harness, so an open-source licence does not remove inference or operating costs.

How Super-agent assigns coding work

A main session handles incoming requests and routes them to workers. Each accepted task becomes a ticket with a defined scope of authority. A worker can follow its issue or pull request, receive subsequent messages and continue the work in its own session.

For example, an authorized developer could ask for a bug fix in a GitHub issue. The worker prepares its checkout, investigates and returns a pull request. A later review comment can reach that worker without starting an unrelated conversation. Messages on related threads can also provide context without granting permission for additional work.

The execution documentation distinguishes active work from work waiting for review, an answer or another event. This makes the panel useful for deciding where a human should intervene, rather than requiring someone to read every agent transcript.

Proactive work is observed by default

The authorization rules separate explicit requests from inferred opportunities. An allowed person can ask the agent to do something. An unaddressed discussion about a bug is treated differently.

The default mode is observe: the system assesses inferred work and records what it would accept, but does not execute it. Switching to auto allows a positive assessment to start work, and still requires at least one matching rule written by the team. An off mode excludes inferred candidates.

Reading a message does not automatically authorize its request. Keeping those concepts separate is important for an agent that receives developer conversation, bot messages and CI output together. Otherwise, a casual suggestion could become a task nobody meant to approve.

The permissions are not a sandbox

Workers run shell commands without approval as the operator's operating-system user. Separate working directories and command wrappers establish defaults; DeepSeek explicitly says they are not a security boundary. A worker can still invoke a program through its absolute path.

The merge controls also need careful reading. Delivery instructions govern agent behavior, while GitHub branch protection must enforce required approvals. The setup guide recommends a dedicated agent account and manual merging initially. Teams should evaluate this in a restricted environment before connecting sensitive repositories or credentials.

This is useful reference material for developers building persistent coding agents. The next practical question is how much review time the system saves after accounting for mistaken task selection and the permissions needed to recover safely. Running more workers only helps if the team can supervise their output without creating another queue of work.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn launches and engineering decisions into clear stories that developers understand and share.

Book a 30-minute call