DeepSeek's 'Super-Agent' bets that coding agents should act before they're asked

DeepSeek has released a prototype of a more autonomous coding agent called 'Super-Agent', Bloomberg reports. The stated goal is to challenge Anthropic's Claude Code. According to a WeChat post on Saturday from the DeepSeek Harness team, the open-source tool can monitor developers' group chats for feedback on coding problems and suggest solutions without being summoned. The available material gives no benchmarks, pricing or release timeline beyond the prototype.

The interesting part is the design choice. Most coding agents today are reactive: a developer asks and the agent acts. A proactive agent that listens to team conversations and steps in on its own is a different kind of product. It could save time by catching problems early. It also needs constant access to internal communications, which is a large expansion of what the agent can see.

Our reading: this fits two trends we have been tracking. First, Chinese labs are competing hard in open-weight AI and are now targeting the agent layer, not just models. Second, the value of enterprise agents increasingly depends on identity, permissions and who can switch them off, not just on how smart the model is. 'Rivals Claude Code' is the framing for now. Until independent tests on hard coding tasks exist, treat it as an aspiration. The open-source release is good news for competition and cost. Whether teams let an agent listen to their chats will depend on how well DeepSeek answers the access-control question.