Key facts
- DeepSeek's Harness team announced Super-agent in a WeChat post on Saturday, November 1, 2026, per BigGo Finance.
- The prototype monitors developers' group chats, identifies coding issues and proposes fixes unprompted, the team said.
- Proactive work is gated by an accept mode; observe is the default and logs assessments without executing them.
- The release is an MIT-licensed TypeScript bundle, version 0.1.0, published in a single commit.
- It runs as a profile on DeepSeek Harness, the company's open-source agent runtime, which is in developer preview.
- Both presets run with danger-full-access as the operator's OS user, and the documentation recommends a dedicated account.
DeepSeek's Harness team announced an open-source coding agent called Super-agent in a WeChat post on Saturday, November 1, 2026, according to BigGo Finance. The unit was formed earlier in 2026 to build agents that can compete with coding products from leading American labs, and the release is its first public demonstration of that effort.
Unlike assistants that wait for a prompt, the prototype is designed to monitor conversations between software engineers, identify coding issues as they surface and offer solutions unprompted, the team said. A user manual posted on GitHub describes the goal as extracting as much value as possible while demanding as little human attention as possible.
The released code is more cautious than that description suggests. Proactive work is gated by an accept mode: off means nothing inferred becomes a candidate, observe is the default and logs a positive assessment as a would-be acceptance without executing it, and auto executes a positive assessment. Requests from people on an allow-list are handled in every mode. The repository states that the agent acts only when asked or when a rule the team wrote allows it.
The release is an MIT-licensed TypeScript bundle, published as version 0.1.0 in a single commit, and runs as a profile on DeepSeek Harness, the company's own open-source agent runtime, which is in developer preview and warns that compatibility-breaking changes will continue. It is not a hosted service: operators run it on their own machines, where it listens to GitHub and, optionally, Feishu, the workplace chat service. Setup needs Node.js 22.19+ or 24+, git 2.32+, Python 3.8+, pnpm and tmux. The bundle brings no model of its own and uses whichever model the Harness profile is set to.
Both presets run with danger-full-access as the operator's OS user, a worker can execute programs by path and read the operator's home directory, and the documentation recommends a dedicated operating-system account holding only the agent's credentials. In AI development, a harness is the software scaffolding wrapped around a large language model that lets an agent coordinate tasks, call external tools and carry out longer workflows. Anthropic has used similar infrastructure to build Claude Code.