Article · AI translation
Thanks to IT Home readers xxy171070, 软媒新友2547452, Wllen, Tokishio, and 阿狗 for the tips!
IT Home reported on October 31 that the official DeepSeek Harness team published an article today titled “Super-agent Design Exploration: Toward a More Autonomous Agent System,” outlining its engineering approach to the Super-agent architecture.
As large models increasingly complete individual tasks without human intervention, the constraint on further productivity gains is shifting toward human attention. Agent systems need to evolve from single-task tools that passively respond to instructions into autonomous collaborators that remain online and handle multiple tasks in parallel.
DeepSeek said mainstream Agent designs still use a single Session as their basic unit, with different Sessions operating independently. When users work on multiple tasks in parallel, they must manually switch between Sessions and maintain each one's context.
According to the DeepSeek team's observations, a person can typically keep track of only 3 to 5 active Sessions at once. Even when a local Agent is extended through CLI, MCP, or external connectors, it may still face synchronous blocking within a single Session, muddled context across tasks, and tasks stopping when the local device goes offline.
They call this new form of Agent and its system architecture a Super-agent, and identify 4 core characteristics: continuity, efficiency, proactivity, and persistence.
Continuity means users can interact with the same Agent through different channels, such as instant messaging and GitHub comments; efficiency means it can handle multiple complex tasks in parallel; proactivity means the system can continuously observe events in its environment and independently advance matters worth addressing; and persistence requires the Agent to remain online and recover from abnormal states during long-running operation.
The overall architecture proposed by DeepSeek consists mainly of 5 parts: a 2-layer architecture, event-driven operation, human-oriented interaction, separation of computation, state, and environment, and a Watchdog.
The 2-layer architecture divides the system into a Router layer and a Worker layer: the Router handles human interaction, maintains global context, and assigns tasks, while Workers carry out specific work concurrently in isolated Sessions. Workers do not reply directly to users; they report status and results to the Router, which handles external communication.
Under the event-driven mechanism, external information such as group chat messages, email and calendar reminders, code review comments, and CI run results is called a Trigger. The system first filters events using program rules, then lets the large model decide whether action is needed. For an event stream associated with an assigned task, a Worker can subscribe to the corresponding Thread, and the program sends subsequent related events directly to that Worker without routing each one through the Router, reducing routing load.
The DeepSeek team also introduced the Inline & Spill and Chain of Recall mechanisms. The former places information likely to be needed soon directly in the model's context window, while keeping content that exceeds the budget or is used infrequently outside the window for the model to read as needed. The latter uses references among objects such as Triggers, Threads, and Worker Sessions to let the model progressively retrieve historical task information by following related IDs.
To reduce human waiting time, the Super-agent also introduces speculative execution, allowing it to begin work in the background based on environmental signals. For example, after receiving a meeting invitation, it can check for scheduling conflicts, gather background information, and prepare a draft reply. In a development collaboration setting, it can try to reproduce a fault and find a fix while an issue is being discussed in a group chat, delivering results to the user only after confirming the problem and finding a simple fix.
For human-Agent interaction, the article proposes an asynchronous ReAct Loop that changes time-consuming tool calls into asynchronous start () and stop () operations. After starting a background task, the Agent can return a Started status without waiting for the task to finish; the background result subsequently enters the Session as a new event.
This allows the foreground conversation to keep receiving new messages from the user. However, the DS team specifically noted that this asynchronous task mechanism has been implemented in DeepSeek Harness but is not enabled by default.
For continuous operation, DeepSeek decouples the Agent Loop's computation, Session state, and execution environment: computation and persistent state reside in the cloud, while the execution environment can remain on the user's local device, with only a lightweight daemon deployed to execute commands and return results. This design supports BYOM (Bring Your Own Machine), connecting the development machine or personal computer a user uses every day to a cloud Agent, as well as connecting cloud-hosted machines or computing environments created on demand.
In addition, the system places an independent Watchdog Session outside the main architecture to form an external control loop. The Watchdog monitors infrastructure issues such as API errors and failed state recovery. It also uses a model to assess whether Workers are making substantive progress and whether their work is drifting from the objective, intervening to fix problems when it detects them.
The DeepSeek team also conducted preliminary experiments comparing the Super-agent, a Super-agent without the direct event delivery mechanism, and a single-Session-plus-Subagent approach under 2 prompt configurations.
Each experiment contained about 300 messages, including about 20 work requests, and each approach was run 5 times. The results showed that the Super-agent's p50 and p95 task completion times were 12.4 minutes and 15.7 minutes, respectively; the single-Session approach with a neutral prompt had a p95 of 29.5 minutes.
The DeepSeek team noted that these experiments used synthetic workloads and a simplified environment. They primarily evaluated the system's scheduling and collaboration capabilities, rather than its ability to execute an individual task in practice.
In a Router layer stress test, as the Trigger arrival rate rose from about 1 to about 40 per minute, the average batch size increased from 1.0 to 2.8 and the time per request rose from 2.5 seconds to 3.9 seconds, while inference time per decision fell from 6.4 seconds to 1.6 seconds. The team concluded that, when events arrive in bursts, the Router can improve effective throughput by combining backlogged events.
The article also points out unresolved issues with the architecture, including how the Router layer scales with the number of tasks, long-term memory spanning months or longer, ways for humans and Agents to collaborate outside Sessions, A2A communication between different Agent instances, and security and privacy protection. In particular, once a cloud Agent connects to a local device, prompt injection, exposure of credentials on the device, and cross-device permission risks all require further work.
The DeepSeek team said an internal Super-agent instance built on these ideas and DeepSeek Harness has been running continuously, taking part in reproducing and fixing issues, responding to code reviews, and iterating on DeepSeek Harness itself.
They also provided a prototype implementation named dsh-superagent-prototype in the article, but explicitly stated that the plugin is intended to validate the technical approach and is not an official DeepSeek product.
References:
dsh-superagent GitHub repository: https://github.com/deepseek-ai/dsh-superagent-prototype