DeepSeek Unveils Open-Source 'Super-Agent' Coding Tool in Challenge to Anthropic's Claude Code

Chinese AI developer DeepSeek has released an open-source prototype of a coding tool that can monitor developers’ group chats, identify problems, and propose fixes without being asked. This move is seen as a direct challenge to Anthropic PBC’s popular Claude Code and other U.S.-based programming tools.

The new system, called Super-agent, was announced by DeepSeek’s Harness team in a WeChat post on Saturday. The team was formed earlier this year with the specific goal of building agents that can compete with leading American coding products.

Unlike conventional assistants, which wait for a prompt and then answer it, the Super-agent prototype is designed to act independently. It monitors conversations between software engineers, identifies coding issues as they arise, and offers solutions without being asked, according to DeepSeek’s team.

The company has published the code publicly on GitHub, allowing outside developers to inspect how it works and test it themselves. This move reflects DeepSeek’s goal of extracting maximum value while demanding minimal human attention.

In AI development, a harness is the software scaffolding wrapped around a large language model that enables an agent to coordinate tasks, call external tools, and carry out complex workflows. Anthropic has used similar infrastructure to build Claude Code, which allows programmers to hand off coding assignments and other development work to AI.

The competition between Chinese and American developers in the field of agents has intensified significantly since OpenAI’s ChatGPT sparked a global AI race. Many companies now see the potential for profit in tools that automate work for engineers and business professionals, leading both sides to rush to develop more capable agents at lower prices.

What sets DeepSeek’s design apart is its approach to engineering. Rather than treating each chat session as a single unit of work, the team has borrowed ideas from classic computer science to keep one agent running continuously across multiple tasks simultaneously.

The architecture splits duties into two layers: a persistent router layer handles interaction with people and decides which task goes where, while isolated worker sessions carry out actual execution. Workers do not interact directly with users; instead, they report back to the router, which responds on their behalf, preventing competing answers from different jobs.

Events drive the system, rather than typed prompts. Messages, code review comments, and CI results are filtered first by deterministic rules and then by a model, ensuring only relevant candidates reach the reasoning stage. The team has also described speculative execution: the agent can attempt a fix in the background and remain silent unless it confirms the problem and finds a workable solution.

Early internal tests suggest that this approach pays off on speed and reliability, although the results come from synthetic workloads and a simplified environment. Completion times measured from request to result message across five runs of each configuration show significant improvements over single-session comparisons.

A separate stress test of the router layer pushed trigger arrival rates from one per minute to about 40. The average batch size rose from 1.0 to 2.8, while time for a single request grew only from 2.5 seconds to 3.9 seconds. This brought inference time allocated to each decision down from 6.4 seconds to 1.6 seconds.

The trade-off is tokens: maintaining multiple sessions costs more cache reads than a single-session setup, resulting in additional compute spent on shorter and more stable completion times. Removing the direct-subscription shortcut left completion times roughly unchanged but increased the router’s workload and token consumption by about 37%.

DeepSeek has flagged four unresolved problems related to its design: scaling the two-layer architecture once a single router can no longer keep up, collaboration with an agent that no longer lives in a chat window, drawing context boundaries as an organization grows, and securing cloud agents that reach into users’ own machines.

The release of Super-agent lands in a crowded field. Anthropic’s Claude Code has established itself as a leading tool for delegating software work to AI, while OpenAI and others are pushing into the same territory. DeepSeek’s decision to publish the prototype rather than ship a finished product suggests it is testing a technical direction in public while converting the design into something commercial.

DeepSeek has raised about 100 billion yuan ($15 billion) in its latest funding round, one of China’s largest, ahead of a planned initial public offering. This capital gives DeepSeek room to continue funding research and agent development against better-resourced American rivals.