DeepSeek has put forward a design for AI coding assistants that can handle more tasks with less human coordination. The Chinese AI company’s DeepSeek Harness team published the paper, and South China Morning Post reported on it Sunday.
Why human attention is the bottleneck
The team’s argument is simple. A single agent’s limit depends on the model. But when several agents work at once, the limit depends on the person watching them. That person can only keep track of so much.
The paper says one person can oversee about three to five active AI agent sessions at the same time. That is a narrow window. Managing more than that means human attention becomes the bottleneck, not model ability.
DeepSeek’s proposal would split tasks across agents and coordinate their work. Humans would still review results and make decisions. The shift is about where people spend their time. Less time routing and monitoring, more time checking output.
A crowded field for agent tools
The report lands in a busy period for AI developers. Anthropic and OpenAI are also pushing agent tools, according to SCMP. The race is not only about model quality. It is also about how many jobs an AI system can run without creating more work for the person in charge.
Crypto traders may have a separate reason to watch Chinese AI models. Cointelegraph reported last October that Chinese models were outperforming US counterparts in a crypto trading experiment, based on CoinGlass data. DeepSeek led that test. It was the only model to post a positive unrealized return, at 9.1%. That result got attention because DeepSeek was built at a fraction of the cost of its US rivals. It does not prove DeepSeek will lead in agent coordination. Trading results and coding agents are different tests, and it is wise to keep them separate.
What to watch next
The paper is a proposal, not a shipped product. The key question is whether the system can coordinate agents without adding hidden human work. If a person still has to check every step, the gain is small. If agents can hand off tasks cleanly, the model of one person overseeing many sessions could become more practical.
For now, DeepSeek is making a case that the next AI bottleneck may be human attention. The company wants its coding agents to work around that limit.
