跪拜 Guibai
← All articles
Artificial Intelligence · DeepSeek · Agent

DeepSeek Super-Agent: An OS-Style Blueprint for Persistent Multi-Task Agents

By 大模型真好玩 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The Thread-subscription design alone cuts token costs by over a third on real workloads, and the Router-Worker split gives a concrete pattern for coordinating multiple tasks without drowning in session sprawl. The prototype is rough, but the architectural ideas transfer directly to production agent systems.

Summary

DeepSeek Harness released Super-Agent, a reference prototype that lifts agent work out of single-session limits. The two-layer design mirrors operating system process scheduling: a persistent Router handles global context and task dispatch without touching tools, while isolated Workers execute with full toolchains and report only results back. Since a human typically juggles three to five active sessions, the architecture lets any channel's messages resolve into the right ongoing task without manual session management.

The system is event-driven with a Fast Path/Slow Path filter: deterministic code screens out noise, and only promising events reach the LLM for semantic decisions. Once a Worker starts a task, that task becomes a Thread, and subsequent events on the Thread bypass the Router entirely. DeepSeek's ablation found that removing this shortcut raised Router decision count by ~30% and read-cache tokens from 23.23M to 31.77M, a ~37% increase. Context stays bounded with Inline & Spill plus chain-of-recall IDs, and storage-compute separation with an external Watchdog keeps the agent alive across restarts.

The source code maps directly to the design: Router logic in src/core, event sources in src/sources, Worker execution in src/executor, and a heartbeat-based health check in src/core/health.ts. Role isolation is enforced at compile time, not by convention. This is a v0.1 reference, resource-hungry and not production-ready, but the architectural patterns, especially the Thread-subscription token savings, are immediately actionable for any team building long-running agents.

Takeaways
— A human can typically follow only 3 to 5 active LLM sessions simultaneously.
— Super-Agent uses a two-layer Router-Worker architecture: Router handles global context and task dispatch without tools, Workers execute with isolated contexts and full toolchains.
— Event triggers are filtered through a Fast Path (deterministic rules) and Slow Path (LLM semantic judgment) to reduce noise and cost.
— Thread subscription direct-pass routes subsequent events on the same task straight to the Worker, bypassing the Router.
— Removing Thread subscription increases Router decisions by ~30% and read-cache tokens from 23.23M to 31.77M, a ~37% token cost increase.
— Context control uses Inline & Spill (recent info in window, long/low-frequency spilled out) and Chain of Recall (unique IDs form a reference graph for old context retrieval).
— Storage-compute separation keeps computation in the cloud and session state in a database, with an external Watchdog for anomaly recovery.
— Source code structure: src/core (Router), src/sources (GitHub/Feishu events), src/executor (Worker), src/core/health.ts (heartbeat-based Watchdog).
— Role isolation is enforced at compile time: Sources only report events, main Session only evaluates/dispatches, Executor only reports facts.
— Super-Agent is a v0.1 reference prototype, doesn't support Windows natively, and consumes significant memory; GitHub setup requires a bot account and a linked repository.
Conclusions

Borrowing the OS kernel scheduler pattern makes sense for agent coordination, but the real cost win comes from the Thread subscription, which cuts a central bottleneck the way a fast path in a kernel avoids unnecessary syscall overhead.

Compile-time role isolation is a more robust security model than runtime permission checks, and it's a pattern worth copying in any agent framework that exposes external event sources.

The prototype's resource hunger and rough edges confirm this is a research artifact, but the numbers from the ablation study give other teams a concrete reason to test the Router-Worker split themselves.

Treating agent state as a database-backed, restartable process instead of a live chat session is what makes 24/7 operation feasible, though it trades away the conversational intimacy most current AI tools rely on.

Concepts & terms
Router-Worker architecture
A two-layer agent design where a persistent Router maintains global context and dispatches tasks, while isolated Worker sessions execute with full toolchains and report back results.
Fast Path / Slow Path
A two-tier filtering mechanism: deterministic code (Fast Path) quickly screens events by source and whitelist, then only candidates go to the LLM (Slow Path) for semantic decision.
Thread subscription
Grouping consecutive events from the same GitHub PR or IM topic into a Thread, and routing subsequent events directly to the assigned Worker instead of passing through the Router, reducing overhead.
Inline & Spill
A context management strategy where recent/frequent info stays in the context window (Inline) and long or low-frequency content is truncated outside (Spill), fetched on demand via tools.
Chain of Recall
Using unique IDs on triggers, threads, and workers to build a reference graph, allowing the model to hop along IDs and retrieve old context from months ago.
Watchdog
An external monitor that watches agent workers for dead loops or off-track behavior and intervenes to correct anomalies.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗