DeepSeek releases MIT-licensed dsh-superagent coding agent

The DeepSeek Harness team released dsh-superagent as a research prototype and made it available under the MIT license[1]. DeepSeek packages it as a bundle for DeepSeek Harness[1]. For teams assessing the release, the license and stated scope are practical starting points: review the code and consider how a one-repository setup fits the work you want to test.

The agent listens to GitHub and can also listen to Feishu[1]. DeepSeek says it converts allowed requests into parallel work items, with each item running in its own agent session[1]. That describes a workflow in which separate tasks can be handled concurrently rather than treated as one continuous session. The available evidence does not specify how many work items can run at once or how teams configure which requests are allowed. Those details matter for evaluating fit, so teams should distinguish the described workflow from limits or controls not specified here.

The release was accompanied by a design post about a more autonomous agent system[1]. DeepSeek says the agent can monitor developer group chats for coding feedback and suggest solutions without being prompted[2]. Separately, reporting describes the proposal as allowing multiple agents to work simultaneously and respond proactively to events without explicit instructions for every task[4]. The proposed system is also intended to reduce the need for people to coordinate multiple agents and monitor their progress[4]. These are useful signals about the intended direction, but the design post and the released prototype should not be assumed to be identical.

DeepSeek says the agent is intended to minimize the human attention it requires while maximizing value[3]. The Straits Times describes a harness as software scaffolding around a large language model that lets agents manage more complex tasks[3]. DeepSeek created its Harness team earlier in 2026 to focus on building AI agents that can take on coding services[3]. Bloomberg reported that DeepSeek’s release is positioned as a challenge to Anthropic’s Claude Code[2]. The launch also comes amid competition between Chinese and US developers to build more advanced and affordable AI tools, according to the Straits Times[3].