DeepSeek Open-Sources Super-Agent Architecture, Tests Show Multi-Task Completion Time Cut in Half
nashnova research
DeepSeek published its first engineering-grade AI Agent architecture, Super-agent, with p95 task completion time dropping from 29.5 to 15.7 minutes in benchmarks — the core idea is letting AI manage multiple task threads on its own instead of waiting for human instructions one by one.
Why aren't today's AI Agents enough?
Individual AI models already handle single tasks well. The bottleneck is human attention — a person can typically track only 3 to 5 active conversations at once.
This means → AI keeps getting faster, but humans can't assign and review fast enough. The person becomes the slowest link in the chain.
Super-agent tackles exactly this: let AI split tasks and run them in parallel, while humans focus on final results.
How does the two-layer architecture work?
Super-agent has two layers. The upper layer, Router, receives messages, assigns tasks, and maintains global context. The lower layer consists of multiple Workers, each executing tasks independently without blocking one another.
In plain terms = the Router is the project manager; the Workers are engineers — the manager assigns work, each engineer runs independently, no one waits.
The system also features a pass-through mechanism: when a Worker picks up a GitHub pull request, subsequent comments bypass the Router and go straight to that Worker. DeepSeek found that disabling pass-through increased Router decisions by ~30% and cache token consumption by ~37%.
What does "speculative execution" mean here?
Super-agent introduces a speculative execution mechanism — when AI spots a potential bug in a group chat, it preemptively tries to reproduce and locate it in the background.
If a viable fix is found, results are submitted. If not, the agent stays silent and does not disturb the user.
This means → AI no longer waits passively for instructions. It anticipates potential issues and attempts fixes — and stays out of the way if it fails.
How does it stay online 24/7 without breaking?
DeepSeek separates compute, session state, and execution environment: compute runs in the cloud, state is saved independently, and the cloud-side agent keeps tasks alive even when local devices go offline.
An independent Watchdog mechanism monitors other agents — if one falls into a futile loop or drifts off target, the Watchdog intervenes automatically.
Message filtering uses a two-tier approach: programmatic rules screen out irrelevant messages first, then the large model decides whether remaining events warrant action, cutting unnecessary token consumption.
How strong are the actual test results?
DeepSeek simulated a collaboration environment with ~300 messages and 20 work requests, comparing Super-agent against a traditional single-session-plus-subagent setup, each configuration tested 5 times.
Super-agent median task completion: 12.4 minutes; p95: 15.7 minutes. The traditional setup hit 29.5 minutes at p95.
In plain terms = for the slowest batch of tasks, Super-agent finished nearly twice as fast.
One caveat: the architecture prototype was just posted to GitHub. Whether these results hold in larger-scale, more complex real-world engineering environments is the key validation milestone the outside world is watching.
市场有风险，内容仅供研究参考，不构成投资建议。