DeepSeek details Super-agent architecture with Router and Worker layers

DeepSeek's Harness team has published a technical article titled “Super-agent design exploration: toward more autonomous agent systems,” describing an architecture it is still developing rather than shipping. The premise is that future agents should not be confined to a single chat window, but should stay online continuously, work on several tasks at once, and act on their own when the outside environment changes.

The team locates the bottleneck in human attention rather than model capability. Most agents still operate with a Session as the basic unit of work, forcing users to switch between windows, remember each task's context and track progress manually; by their observation, a person can typically follow only three to five active sessions at a time. Connecting agents to more platforms through CLIs and MCP does not solve serial execution inside one session, tangled multi-task context, or tasks that stall when a local device goes offline.

To address that, DeepSeek proposes a two-layer structure of Router and Worker. The Router runs continuously as a global coordinator, receiving input from chat, GitHub, email and other platforms, deciding how work should be distributed and maintaining context across tasks. Workers have their own context and tool access, handling reproduction, code review or feature development, and can call further subagents when needed. For the user, it always looks like one assistant; behind it, several independent agents may be working at once. A subscription mechanism reduces coordination overhead by routing later comments and status changes on a pull request straight to the Worker that picked it up.

Proactivity is the more interesting half of the design. Chat messages, email, calendar reminders, GitHub comments and CI results all become Triggers, filtered in two stages: deterministic rules discard irrelevant chatter, then a model judges whether the remaining context warrants action. DeepSeek borrows the systems idea of speculative execution, running a “silently reproduce and fix” policy in internal user groups. The agent reproduces a reported bug and looks for a fix in the background, surfacing results only when the problem is confirmed and the remedy is relatively simple; otherwise it stays quiet. In one demonstration, a fix pull request was already open while the team was still debating the issue in chat.

On responsiveness, the team describes an asynchronous ReAct Loop that moves slow tool calls to the background so the foreground session keeps accepting instructions, including requests to change direction or stop the task. The mechanism is implemented in DeepSeek Harness but not enabled by default. DeepSeek also separates compute from storage: the agent's execution loop runs in the cloud, session state lives in a database, and the repository, files and toolchain can stay on the user's device behind a lightweight daemon, so work resumes after a machine reconnects. A separate Watchdog checks whether tasks are making progress or looping unproductively.

A preliminary experiment simulated streams of chat and GitHub CI messages, about 300 messages per run with roughly 20 genuine work requests, repeating each configuration five times. Super-agent completed tasks with a p50 of 12.4 minutes and a p95 of 15.7 minutes, against a p95 of 29.5 minutes for a conventional single-session setup given a neutral prompt — which, in all five runs, never invoked a subagent at all, leaving every task in the main session where new messages kept interrupting it.

The gains come at a cost. Removing the direct event-passing shortcut raised the number of decisions the Router had to make by about 30 percent, and total cache-read volume grew from 23.23 million to 31.77 million tokens, roughly 37 percent. When the event arrival rate rose from about one per minute to 40, batching kept per-request handling time to a rise from 2.5 to 3.9 seconds. The authors caution that the test used synthetic messages and simulated slow tools, so it measures scheduling and collaboration rather than real coding quality.

Part of the design now exists as the dsh-superagent-prototype plugin, but DeepSeek frames the work as engineering exploration, not a standalone product launch. Long-term memory across weeks or months, security and privacy — prompt injection or a server breach could expose repositories, sensitive data and local credentials — and the question of how users inspect progress, intervene and decide which actions need human approval all remain open. The team concedes it introduced no new theory, applying classic systems ideas such as task scheduling, cache management, event-driven design and compute-storage separation to agents; what is actually being redefined is how many AIs one person can supervise.