Chinese AI company DeepSeek has outlined a new architectural approach for AI agents that aims to reduce the amount of direct human coordination required when managing multiple tasks.
The proposal, shared by the company’s DeepSeek Harness team, focuses on overcoming a practical limit in current agent systems: while models have grown stronger at completing individual jobs, people can typically monitor only three to five active sessions at once.
Existing agents usually treat each conversation or session as a separate unit of work.
Users must switch between them, track context manually, and intervene frequently.
The Harness team argues that single-task performance is constrained mainly by model capability, whereas multitasking is constrained by human attention.
Their proposed “Super-agent” seeks to treat the agent as a continuous entity rather than a collection of isolated sessions.
The design emphasizes four qualities: continuity (so users interact with what feels like one agent across messaging apps and code platforms), efficiency (by running complex work in parallel), proactivity (by spotting issues that need attention without an explicit command), and persistence (by remaining available over long periods and recovering from interruptions).
In practice, this could allow the same development task to move between a group chat and a GitHub discussion while the system maintains shared context.
At the core is a two-layer structure.
A resident Router handles user communication, keeps a shared understanding of ongoing work, and assigns tasks.
Multiple Workers then execute those assignments independently in isolated environments and report results back through the Router.
Supporting elements include event-driven triggers, asynchronous interactions, and an external monitoring component that helps manage routine coordination.
The intent is to free people for higher-level review and decisions rather than constant session management.
An internal version of the system is already in use at DeepSeek.
It assists with reproducing and fixing software issues, responding to code reviews, and improving the Harness framework itself.
For instance, if developers discuss a possible bug in a group chat, the agent can attempt to reproduce the problem and propose a solution in the background without being directly addressed.
The team has also released a research prototype as open-source code under the MIT license.
The public version, packaged for use with DeepSeek Harness and focused on a single GitHub repository (with optional chat integration), is explicitly described as exploratory reference material rather than an official product.
It differs from the internal system, and some features described in the design document remain under development.
In simulated tests, the approach completed tasks faster than a conventional single-session setup, though the evaluation measured scheduling efficiency rather than final code quality.
The work arrives as Chinese and U.S. companies intensify competition around practical AI agents for coding and other professional tasks.
DeepSeek’s Harness framework, introduced earlier as an open-source tool for connecting models to external tools and workflows, provides the foundation for these experiments.
By shifting emphasis from pure model performance toward reliable multi-task coordination and lower human overhead, the Super-agent concept reflects a broader industry move toward systems that can operate more independently while still leaving final judgment with people.
