Ads_970x250

DeepSeek Releases Autonomous Coding Agent Prototype to Rival Claude Code

The open-source system monitors developer requests, assigns coding tasks and returns results with limited human intervention.

Topics

  • Chinese AI startup DeepSeek has released an open-source prototype of an autonomous coding agent designed to identify development tasks and execute them with less human supervision, stepping up competition with tools such as Anthropic’s Claude Code.

    Developed by DeepSeek’s Harness team, the system, called dsh-superagent, monitors activity in GitHub repositories and, optionally, workplace conversations on the Feishu messaging platform.

    It can identify authorized requests, assign coding tasks and return results, including proposed code changes, without requiring developers to initiate and supervise each session individually, according to DeepSeek’s Harness team on X.

    The team published the prototype on GitHub on 31 October, describing it as a research project rather than a production-ready product.

    The system uses a two-layer architecture comprising a Router, which evaluates incoming requests and assigns tasks, and Workers, which execute them in separate coding sessions.

    It also uses conventional software rules to process routine events before involving a large language model, reducing unnecessary AI processing.

    The approach is intended to address a limitation of existing coding assistants, which typically require developers to initiate tasks, monitor progress and manage multiple sessions.

    DeepSeek said its objective was to “take as little human attention as possible, and give as much value as possible.”

    Unlike conventional coding assistants, the prototype can operate continuously, although autonomous task execution requires explicit authorization or rules established by the development team.

    The company cautioned that the published code differs from its internal implementation and that some features described in its accompanying research remain under development.

    The prototype is built on DeepSeek Harness, an open-source framework that provides the tools and infrastructure needed for AI models to execute tasks and manage development workflows.

    A harness refers to the software infrastructure built around a large language model (LLM) that allows it to execute complex tasks, manage workflows, and operate with greater autonomy. 

    Companies such as Anthropic are increasingly investing in these systems to make AI agents more effective at coding and other professional tasks.

    The particular prototype has drawn attention for its underlying architecture. One X user, highlighted the system’s approach to managing AI sessions, arguing that sessions should be treated as low-level system components rather than adding to the cognitive burden on developers.

    The user pointed to its two-layer Router/Worker architecture, which separates task coordination from execution, and its fast/slow processing paths, which use deterministic code to filter routine events before involving an LLM. The design also separates cloud-based orchestration from local execution, potentially helping keep sensitive credentials within internal networks.

    Its release comes as AI developers compete to build coding agents capable of handling increasingly complex software projects with less direct supervision.

    DeepSeek has also secured substantial funding to support its expansion. Bloomberg reported on 22 October that the company had raised about $15 billion from investors, including Tencent and battery manufacturer CATL.

    The Information subsequently reported that DeepSeek had closed the funding round at about $15.9 billion and was preparing for a potential listing on Shanghai’s STAR Market.

    The investment gives DeepSeek additional resources to compete with US AI developers, including Anthropic and OpenAI, as attention shifts from model performance toward agents capable of completing work independently.

    Topics

    More Like This

    You must to post a comment.

    First time here? : Comment on articles and get access to many more articles.