DeepSeek's Harness team published a design essay on October 31, 2026 titled Exploring Super-agent Design: Towards a More Autonomous Agent System. The piece sketches a Router and Worker two-layer, event-driven architecture for agents that stay online and advance many tasks at once. The same day, deepseek-ai opened dsh-superagent-prototype on GitHub under the MIT license. The README calls it a research prototype for reference only, not an official DeepSeek product, and says it is not identical to the internal build or the essay.

Key Facts

Details
Who DeepSeek Harness team
Essay Exploring Super-agent Design: Towards a More Autonomous Agent System (Chinese WeChat post, Oct 31, 2026; cited by the official README and secondary reports)
Core design Resident Router for human interaction and dispatch; isolated Worker sessions for concurrent execution
Target traits Continuous, efficient, proactive, persistent
Async tools start() / stop() already implemented in DeepSeek Harness, not enabled by default
Synthetic test ~300 messages per run (~20 work requests), 5 runs; Super-agent completion p50/p95 = 12.4 / 15.7 min; neutral single-session + Subagent p95 = 29.5 min
Open source dsh-superagent-prototype, MIT; research reference, not a shipped product
Internal status Team says Super-agent instances built on these ideas already run continuously

Background

Mainstream agents still treat one Session as the unit of work. Sessions do not share context, so people who juggle several tasks switch windows and keep each thread's state in their heads. The essay says a person can usually follow only about 3 to 5 active sessions. Extending a local agent with CLI, MCP, or connectors does not remove synchronous blocking inside one session, mixed context across tasks, or the stop when the local device goes offline.

Super-agent is the name DeepSeek gives to a system meant to stay continuous across channels (IM, GitHub comments, and similar), run several complex jobs in parallel, watch environment events and push work on its own when rules allow, and keep running with recovery after failures.

Router and Worker. The Router is a resident session that takes human input from every entry point, keeps cross-task context, and dispatches work. It does not do the heavy execution itself and may run a lighter model without Bash. Workers hold isolated context, mount full tools, and usually take the strongest model. Workers do not talk to users. They report status to the Router, which replies outward. Inside one hard task, a Worker may still open Subagents; that stays invisible to the Router. The essay separates this from a typical Agent Team, where a lead decomposes one task and coordinates members closely. The Super-agent Router only tracks relationships among separate tasks.

Events and Threads. External signals such as group messages, email and calendar reminders, review comments, and CI results are Triggers. A programmatic filter runs first; an LLM then judges what needs action. After a Worker is assigned, it can subscribe to a Thread so later related events go straight to that Worker without another Router hop.

Other mechanisms named in the essay. Inline & Spill treat the context window like a cache. Chain of Recall walks object IDs across Triggers, Threads, and Worker sessions. Speculative execution can start background work from environment signals and speak only when a result is ready. For long uptime, compute and session state sit in the cloud while a thin local daemon executes commands (BYOM: Bring Your Own Machine). A separate Watchdog session watches infrastructure faults and stalled Workers.

The open prototype is packaged as a DeepSeek Harness bundle for one GitHub repository, with optional Feishu. Docs such as 000 Philosophy state the operating rule: take as little human attention as possible, and give as much value as possible. The README warns that much of the code was written by coding agents, that it may contain bugs, and that some designs in the essay are not implemented in the public tree.

What We Don't Know Yet

  • No stable public URL for the original WeChat essay was found at publish time; details above rest on the official GitHub README plus matching figures in IT之家 and AgentsPulse.
  • How large the internal Super-agent fleet is, which teams use it, and how often humans still override Router decisions.
  • Whether the published p50/p95 times hold on real coding workloads; the authors label the load synthetic and the results preliminary, and coding was simulated by a time-consuming tool.
  • When, if ever, Router/Worker, thread subscription, or async start()/stop() become default DeepSeek Harness product features.
  • Concrete mitigations for prompt injection, endpoint credential exposure, and cross-device privilege when a cloud Agent Loop can drive local machines.

Sources