DeepSeek's “Super-Agent” takes on Claude Code with an agent that acts unprompted, which brings new security questions

DeepSeek has released a prototype of a more autonomous coding agent called “Super-Agent”, according to Bloomberg. The aim is to compete with Claude Code, Anthropic's popular coding agent. The tool is open source. According to a WeChat post on Saturday from DeepSeek's Harness team, the agent can proactively monitor developers' group chats, spot feedback about coding problems and suggest solutions without being called on. The available information stops there: there are no benchmarks, no comparative data and no release details beyond the fact that it's a prototype.

The interesting part is the design, not the competitive positioning. Most coding agents wait for an instruction. This one listens to the conversation and steps in on its own. That changes how people work with it. The agent stops being a tool someone calls up and becomes a participant in the team.

Our thesis: it fits a trend we've been tracking. China is closing the gap fastest in open-weight models, and DeepSeek is now pushing into the agent layer as well. Still, a prototype aimed at rivaling a product is not a product that rivals it. We'll judge it on independent evaluations, not on stated goals.

Our reading: an agent that reads group chats without being asked raises exactly the issues we've flagged in agent security. It has access to a channel where people share context, sometimes credentials, and sometimes untrusted messages that could be used to manipulate it. The relevant question is not how good its suggestions are. It's what permissions the agent has, which channels it can read and act on, and how its sequence of actions is monitored. Proactive agents can greatly speed up software development, and the fact that this one is open source makes the technology more widely available. To earn trust, though, autonomy will need clear limits built into the architecture.