DeepSeek Harness Team Publishes Super-agent Architecture Design and Prototype

The DeepSeek Harness team published an engineering design exploration for a more autonomous super-agent system, arguing that single-task agents are strong enough and the bottleneck is how many sessions a person can monitor. The design replaces a single chat window with a two-layer Router/Worker architecture, event-driven triggers with filtering and direct pass-through, asynchronous ReAct, separation of compute, state and execution environment, and an external Watchdog. A research prototype repository, dsh-superagent-prototype, accompanies the article and is described as reference-only code, not an official DeepSeek product.
Event version 1.
Key insights
Bottleneck shifts to human attention
Single-task coding agents are considered strong enough, and the limiting factor becomes how many active sessions one person can track, typically about 3 to 5. The design targets multi-task throughput rather than a single-task ceiling.
Two-layer Router and Worker split
A persistent Router layer handles human interaction, global coordination and memory without executing tasks, while context-isolated Workers run single tasks with full tooling and report back through the Router. The Router can use a lighter model because it only dispatches between tasks.
Event-driven triggers with direct pass-through
Inputs become semi-structured triggers from group chats, email, calendars, code review and CI, filtered first by deterministic rules and then by an LLM. Once a Worker subscribes to a thread, subsequent events pass directly to it, reducing Router decisions by about 30 percent and read-cache tokens by about 37 percent.
Compute, state and environment separated
The agent loop runs in stateless elastic cloud containers with session state in a database, while a lightweight daemon on the local device executes commands against the heavy execution environment. This keeps the agent online persistently and aligns with the brain-hand separation concept.
Prototype results and open problems
On a synthetic load of about 300 messages and 20 work requests, super-agent task completion times were p50 12.4 minutes and p95 15.7 minutes, better than a single-session plus subagent setup. Open issues include Router partitioning, long-range memory, task presentation after sessions are hidden, and security boundaries for cloud loops with local daemons.