DeepSeek Unveils an Autonomous Coding Agent That Works Without Waiting for Instructions

DeepSeek has published an open-source prototype for a coding agent that keeps working across group chats, code repositories and other developer tools without waiting for a person to tell it what to do next, a design the Chinese company says sets its system apart from the coding agents now sold by Anthropic and OpenAI.

The system, called a “Super-Agent” and built by DeepSeek’s Harness team, splits the work between a Router layer and a group of Worker sessions. The Router holds the wider context of a task and handles communication, while each Worker carries out a separate coding job in parallel, according to a company post on Weibo reported by Stocktwits. That structure lets one agent follow a task as it moves from one channel to another: DeepSeek said an internal version can track a development issue from an instant-messaging conversation into GitHub while also watching public developer discussions for new information.

The difference from today’s tools, DeepSeek said, is autonomy. Most coding agents work inside a single session: a developer gives an instruction, the agent does the work, then waits for the next command. DeepSeek said that setup makes human attention, rather than what the model can do, the limit when several tasks run at once, since a person can typically supervise only three to five active agent sessions at a time.

An agent that acts on its own

The prototype is built to act on events instead of waiting for orders. If someone reports a bug in a group chat, the agent can spot the report, try to reproduce the problem and, if it finds a fix, submit a proposed code change for human review, without being tagged or called on by name. If it fails, it stays quiet rather than sending a notice nobody asked for, according to a summary of the design by Interesting Engineering.

In an internal test described by the company, the approach finished one task in just under 16 minutes, compared with more than 29 minutes for a conventional single-session approach. Interesting Engineering cautioned that one timing result on one task does not show the system beats competing coding agents, and that the technical paper gave too little detail to judge how widely the result applies, how task quality was measured, or whether it holds across different kinds of software work.

A research prototype, not a product

DeepSeek published the code on GitHub under an MIT license and described it as a research prototype for reference only, not an official DeepSeek product. Some of the designs in the accompanying paper are plans for future work and are not in the published code, and security, privacy and reliability remain open problems, according to the same summary.

The release lands as Anthropic prepares for an expected stock market listing: the company filed its registration statement with the U.S. Securities and Exchange Commission in late October and is expected to start its IPO roadshow later this month, Stocktwits reported, citing a Financial Times report that Blackstone does not expect Anthropic to secure an investment-grade rating right after going public.

Reporting based on coverage by Stocktwits and Let’s Data Science.