# DeepSeek's Super-agent Design - Continuous, Proactive, Persistent AI Agents
DeepSeek Harness, the open-source "everything is a plugin" agent framework, just published an internal design document about what they call the Super-agent: an agent architecture that is continuous, productive, proactive, and persistent.
Their internal instance already runs across IM and GitHub, monitoring issues and opening PRs without being told.
The four characteristics
The article argues that current agent systems have a fundamental problem. A task started in one session can't be seen by another. Each session is a fresh start. The bottleneck isn't model capability anymore - it's human attention.
The solution is an agent decoupled from any single manually-driven session:
| Characteristic | What it means | |---|---| | Continuous | Same agent across IM, GitHub, CLI. No more "does session X know about Y?" | | Productive | Multiple complex tasks in parallel. Takes the next instruction before the previous one finishes | | Proactive | Monitors events in its environment. Acts on what matters, stays quiet otherwise | | Persistent | Always online. Recovers from anomalies automatically |
The five-module architecture
1. Two-level architecture - one level for continuous interaction (chat, PR comments), one for deep work (long-running tasks). They share context but don't block each other.
2. Event-driven design - the agent subscribes to events (new GitHub issue, CI failure) and triggers actions. This is what makes it proactive rather than simply reactive.
3. Full-duplex multi-channel interaction - instructions on one channel (IM), results on another (GitHub). New instructions arrive before the first task finishes.
4. Compute-state-environment separation - the agent's state lives independently of where it's running. A task can start on one machine and finish on another.
5. Watchdog - an external monitor that detects stuck tasks, crashes, and resource leaks. The agent heals itself.
> "Rather than proposing new theory, this design applies classic engineering ideas from computer systems to building agent systems."
Already running inside DeepSeek
The article is not theoretical. The team's internal Super-agent instance, built on DeepSeek Harness, already:
- Reproduces and fixes bugs across IM and GitHub
- Responds to code reviews
- Monitors public group chats and opens PRs unprompted
- Develops DeepSeek Harness itself every day
The async task mechanism is in DSH code but not yet enabled by default. The superagent prototype plugin is experimental. But the architectural direction is clear.
What this means for anyone building agent systems
The design applies well-known systems engineering patterns - event-driven architecture, state separation, watchdog monitors - to the agent problem. It is not proposing new theory, it is applying what we already know about building reliable distributed systems.
For anyone running an agent harness today, the relevant question is not "should I build this?" - it is "which of these patterns can I adopt first?" The compute-state separation alone (characteristic 4) would solve the problem of a Hermes session not knowing what Claude Code did, and vice versa.
The full article is on X (in Chinese, machine translations available). DeepSeek Harness is open source on GitHub under MIT.
https://github.com/deepseek-ai/deepseek-harness