DeepSeek Unveils Autonomous Coding Super Agent Architecture to Rival Anthropic and OpenAI
Chinese artificial intelligence enterprise DeepSeek has introduced a framework for a more autonomous coding assistant requiring minimal human oversight, according to the South China Morning Post. The corporate initiative expands competition against international rivals OpenAI and Anthropic in the rapidly growing market for automated programming software. The proposed architecture, designated super agent, seeks to eliminate the necessity for human operators to coordinate multiple separate assistant sessions. Technical specialists detailed the development within an engineering paper published by the company's DeepSeek Harness engineering division on Saturday.
Human Attention Bottlenecks and Centralized Task Delegation
The technical publication noted that while individual task boundaries depend on underlying model capabilities, multitasking constraints are dictated by available human attention. DeepSeek researchers determined that an individual engineer can realistically oversee only three to five simultaneous assistant sessions without experiencing productivity bottlenecks. The new system addresses this administrative constraint by automatically distributing complex assignments among subsidiary agents and managing their execution, the South China Morning Post reported. Under the proposed framework, a central coordinating agent assigns responsibilities to independent worker units that execute programming tasks in parallel.
By default, the prototype restricts automatic merging to approved pull requests involving tickets with fewer than 200 modified lines of code, while never activating auto-merge through the agent account itself, according to AICoder.
Benchmark Performance and Proactive Chat Monitoring
Worker agents operate concurrently and transmit status reports to the coordinating unit, diminishing the requirement for constant human supervision. As a practical application, researchers explained that an autonomous agent could identify a reported defect in developer messages and attempt to reproduce the fault independently, the South China Morning Post noted. When a viable bug fix is verified, the agent submits proposed code modifications for human review while remaining silent if unresolved to avoid distracting engineering personnel. In experimental benchmarking trials, the autonomous multi-agent method completed a benchmark software task in under 16 minutes, compared to over 29 minutes under conventional manual workflows.
Harness Engineering Framework and Capital Backing
DeepSeek deployed an open-source prototype of the autonomous programming tool designed to challenge Anthropic's Claude Code, Moneycontrol reported. The open-source program can observe developer chat channels for reported code defects and propose functional resolutions without explicit user invocation. The software aims to minimize required human attention while delivering maximal operational utility, according to documentation published on GitHub, Moneycontrol noted. DeepSeek established its dedicated Harness division earlier this year to construct specialized agents capable of matching developer services from leading international laboratories. Software harnesses function as structural scaffolding surrounding large language models to enable automated systems to execute sophisticated workflows. To support its development roadmap against competitors, DeepSeek secured approximately 100 billion yuan, equivalent to 15 billion dollars, in its latest financing round ahead of a public listing.