On October 31, 2026, DeepSeek's Harness team announced Super-agent, an open-source coding agent prototype that watches a team's GitHub repository and optionally Feishu (Lark) group chats, then steps in with fixes without being prompted for every task.

The code was published under the MIT license, and Bloomberg framed the move as a challenge to Anthropic's Claude Code.

DeepSeek is careful about labels: it calls the code a research prototype for reference only, "not an official DeepSeek product," and warns it is not production-ready and differs from the version the team uses internally.

The release follows a striking run for DeepSeek's open-source agent tooling. According to Startup Fortune, the deepseek-harness repository hit 177,904 GitHub stars within eight days of its August 13, 2026 launch and has since climbed past 188,000 under the MIT license.

shiftdelete.net reports the separate dsh-superagent repo passed 4,000 stars within a day of going public.

That speed matters because the plumbing is open: developers can clone, inspect, modify, and self-host the code while choosing any underlying model supplier, which puts pressure on agent-framework startups building orchestration layers on top of commercial models.

Architecturally, Super-agent is built as a bundle for DeepSeek Harness, whose plugin runtime is called Cordis, a framework for composing components and reversing their effects when unloaded.

The prototype splits the agent into two layers: a persistent Router talks to people and keeps shared context, while isolated Worker sessions run tasks in parallel, so a new request can start before the last one finishes.

A Worker that owns a pull request keeps following its comments directly, and a Watchdog process watches for failures and stalled loops.

Work items run in their own session and directory before the agent replies in the original thread, often with a pull request, and a local web panel lists only the items needing a human decision.

The project's stated rule is to "take as little human attention as possible, and give as much value as possible," but DeepSeek is explicit about constraints.

Setup requires a dedicated GitHub user account for the agent, Node.js 22.19 or later, GitHub CLI 2.48 or later, and a model provider configured in Harness; only github.com is supported, not GitHub Enterprise Server.

Only github.com is supported, and proactive work nobody asked for starts in a mode that merely records what the agent would have done.

DeepSeek warns the software has not undergone a security audit, its sandboxing and approval controls do not guarantee isolation, and sessions run with full access as the operator's OS user, recommending a separate OS account.

shiftdelete.net ties that warning to recent supply-chain scares such as a tampered Claude Code extension that planted an SSH backdoor.

The release separates DeepSeek's open-source agent tooling from its model API business and signals how Chinese and US labs are now competing on agents, not just chatbots.

Bloomberg reported DeepSeek created its Harness team earlier this year to build agents that can compete with US coding services, and has raised about 100 billion yuan ($15 billion) ahead of a planned IPO.

The competitive backdrop is heated: Anthropic has separately accused DeepSeek of large-scale "distillation attacks" against Claude, an allegation documented in February 2026, and shiftdelete.net reports US labs including the GPT-6 Astra distillation report have named DeepSeek in copying claims.

For now, Super-agent is a blueprint developers can study and test, with the ecosystem invited to build the next ones.