Jack Paul 
DeepSeek has outlined a design for AI coding assistants that can handle more tasks on their own. The goal is to cut the amount of human work needed to coordinate and watch over several agents at once. The company’s DeepSeek Harness team published the paper, and Singapore’s South China Morning Post reported on it Sunday.
Why one person cannot watch too many agents
The team points to a practical limit. A person can keep track of only three to five active AI agent sessions at the same time, according to the paper. That limit is not just about model skill. The researchers note that single-task limits depend on model capabilities, while multitask limits depend on human attention. In other words, the bottleneck can be the person in the loop, not the software alone.
DeepSeek’s proposal would spread tasks across agents and have the system coordinate their work. Humans would then review results and make decisions, instead of stepping into every handoff. That sounds simple enough, but coordination is hard. Agents can duplicate work, miss context, or drift from the plan. The paper’s design is meant to reduce that load, though it does not remove human review.
A crowded field for coding tools
The report comes as AI developers push harder into coding assistants and multi-agent setups. US companies Anthropic and OpenAI are among the competitors, SCMP noted. DeepSeek’s paper is a design proposal, not a finished product, so the real test will come when such systems are used on messy projects with unclear goals. That is where oversight tends to matter most.
Crypto traders already noticed DeepSeek
DeepSeek has also drawn attention outside coding. Cointelegraph reported last October that Chinese AI models were beating US counterparts in a crypto trading experiment, based on data from blockchain analytics platform CoinGlass. DeepSeek led the test and was the only model to post a positive unrealized return of 9.1%. The result surprised some traders, especially since DeepSeek was built at a fraction of the cost of its US rivals.
That trading result does not prove DeepSeek’s new agent design will work. It does show why its research gets watched. For now, the paper makes a modest bet: if agents can share the coordination work, people can spend less time monitoring sessions and more time judging outcomes. Perhaps that is the useful shift. Less supervision does not mean no supervision.