DeepSeek's Super-Agent Architecture Finds Bugs and Files PRs Without Human Hand-Holding

Background

Developers' daily work involves switching between multiple platforms: discussing solutions in IM group chats, submitting code to GitHub, and revising implementations based on PR comments. Traditional coding agents keep these operations in separate sessions, forcing users to manually maintain context and relay task progress between sessions. DeepSeek's team observed that a person can usually only track 3 to 5 active sessions at once, which limits multi-task collaboration efficiency. As models get stronger at single tasks, human attention becomes the bottleneck.

Super-agent solves exactly this problem. Users face one agent that manages multiple underlying sessions from group chat decisions to GitHub PR reviews. In DeepSeek's internal examples, the agent understands that adopting Plan A in a group chat and switching to Plan B in a GitHub comment belong to the same effort, and adjusts execution direction accordingly. The agent also introduces speculative execution: while a bug is discussed in a group chat, the background starts reproducing and locating it; when it finds a viable fix, it hands it to the user; if no effective result is found, it stays quiet.

Core Principles

Super-agent uses a two-layer architecture. The upper-layer Router coordinates the global picture, responsible for receiving messages, assigning tasks, and maintaining context. The lower-layer Worker is a set of agent sessions that execute tasks; multiple tasks can run concurrently. The key design is a subscription and direct-delivery mechanism: once a Worker takes over a GitHub PR, it directly subscribes to subsequent comments, so new review comments bypass the Router and go straight to the corresponding Worker. Old tasks continue to be followed up, and new tasks are assigned promptly.

Message filtering uses a two-tier mechanism: programmatic rules first filter out irrelevant messages, then a large model judges whether remaining events need action, reducing token consumption. The async ReAct Loop lets the AI respond to new instructions while a time-consuming task is still running, without waiting for the previous task to finish. This mechanism is implemented in DeepSeek Harness but is not enabled by default.

24-hour availability depends on separating compute, state, and execution environment. Agent computation runs in the cloud, session state is stored independently, and the local device keeps the working environment. When the computer goes offline, the cloud agent maintains task state and resumes execution when the device is back online. The Watchdog mechanism acts as an independent supervising AI that both detects program failures and judges whether a task is stuck in a pointless loop or drifting from its goal, and tries to intervene and fix it.

Limitations

The efficiency gain comes with token overhead. After canceling the event direct-delivery mechanism, Router decisions increase by about 30% and cached-token reads by about 37%. Reasonable event routing cuts operating costs, but the multi-layer architecture itself introduces extra consumption.

Initial experiments ran in a simulated environment of about 300 messages and 20 work requests, with each configuration repeated five times. Super-agent median completion time was 12.4 minutes, 95th percentile 15.7 minutes; under neutral prompts, the traditional single-session approach hit a 95th percentile of 29.5 minutes. The experiment is small, and performance in more complex scenarios remains to be verified. The async ReAct Loop and speculative execution are not enabled by default, so their maturity needs further observation.

What It Means for Developers

Super-agent frees agents from single-session limits and enables continuous cross-platform collaboration. Developers no longer need to restate context across multiple sessions; the agent can track the full flow from group chat discussion to code commit. Proactive response and speculative execution cut wait time; the agent tries solutions in the background and presents results when found without interrupting otherwise.

Parallel execution advances multiple complex tasks at once, avoiding the situation where one time-consuming task blocks later work. 24-hour availability and automatic recovery support long-running operations, with attempts to self-heal when anomalies occur. These features are moving agents from 'waiting for a human to assign each task' toward 'proactively discovering problems and pushing work forward.'

DeepSeek opens up the architecture ideas, also providing a reference for developers exploring similar systems. The GitHub repository already has prototype code, which developers can build on for further experiments and improvements.

Original URL: https://tools.wflynn.cn/aitools/trend-radar

Follow the official account 'YourTools' for daily GitHub trends and open-source tool briefings: