DeepSeek’s Harness team has released a technical paper describing a “super agent” architecture aimed at coordinating multiple AI agents so they can divide work automatically and handle tasks in parallel, reducing the need for humans to supervise each step one by one.
The paper argues that while the ceiling for a single task is set by model capability, the ceiling for multitasking is set by human attention. DeepSeek estimates that one person can effectively keep track of only three to five active agent sessions at the same time, which it identifies as a core constraint on scaling productivity gains from multi-agent deployments.
A central coordinator assigns work across independent agents
DeepSeek’s proposed solution uses one central coordinating agent to distribute tasks to multiple independent agents. Those agents run in parallel and report progress back to the coordinator, while the human role shifts to decision-making and reviewing results.
In the architecture outlined in the paper, agents can also respond to events on their own, such as incoming messages or code updates, instead of waiting for explicit instructions before every action.
Paper example shows an agent reacting to a bug report on its own
The paper gives an example drawn from a typical software development workflow. If an agent detects a reported software bug in a group conversation, it can attempt to reproduce and fix the issue even without being directly instructed to do so. If it finds a workable solution, it submits a code change for human review. If not, it stays silent to avoid unnecessary interruptions.
Test data showed task completion time fell from more than 29 minutes to under 16
According to the team’s experimental data, the same task was completed in less than 16 minutes under this approach, compared with more than 29 minutes using a conventional workflow, nearly halving the time required.
The proposal arrives as AI developers compete to build more autonomous coding assistants. The report notes that Anthropic’s Claude Code and OpenAI’s Codex have recently been able to handle increasingly complex software development tasks. DeepSeek is trying to push automation one step further by having multiple agents work at the same time and react proactively.
The team has also released an open-source research prototype of the super agent on GitHub, while stressing that the work remains an architectural exploration rather than a mature product.
Security, privacy and reliability remain open questions
DeepSeek also acknowledged security and privacy risks in the architecture. One direct concern raised by the team is that malicious instructions could use agents connected to other services or systems to launch attacks and obtain permissions they should not have. The more connections agents have and the wider their reach, the larger the potential attack surface becomes.
The industry is also not settled on whether faster automatically means better. Liu Ming, an AI entrepreneur in Beijing, wrote on Zhihu that the real challenge in AI software development today is ensuring “reliability,” not simply making tasks run faster. That criticism points to a practical issue in agent collaboration: if higher speed comes with a higher error rate, the real value to enterprises is limited.
The report also links the model to crypto-AI experimentation
The article says multi-agent collaboration also has implications for the intersection of crypto and AI. According to an autonomous trading experiment conducted by on-chain data platform CoinGlass in October 2025, DeepSeek was the only model in the test to post a positive return, with an unrealized return of 9.1%. Its training cost was about $5.3 million, lower than that of U.S. peers.
In that experiment, each model was allocated $10,000 and traded on the decentralized exchange Hyperliquid. The report says that if agents can divide work and make decisions in parallel on their own, the scale and complexity of such experiments could move to a new level.
What to watch next
The article says the key issue for development teams and crypto teams is not only which model runs faster, but how security boundaries and responsibility should be defined once agents begin taking over coordination work on their own. By open-sourcing the prototype, DeepSeek has effectively pushed that question back to the broader ecosystem.
Two follow-up points stand out. One is whether the open-source prototype will be adopted by other agent frameworks and evolve into a de facto standard. The other is whether Anthropic and OpenAI will introduce similar “central coordinating agent” designs, shifting competition from single-model capability toward multi-agent scheduling.


