---
id: "lib-2026-10-31-deepseek-agent-deepseek-open-sources-its-super"
type: "library"
title: "刚刚，DeepSeek重磅开源：Agent底座架构！/ DeepSeek open-sources its Super-agent architecture"
lang: "ZH"
org: "量子位 QbitAI"
source_type: "Practitioner post"
date: "2026-10-31"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-31-deepseek-agent-deepseek-open-sources-its-super/"
original_url: "https://www.qbitai.com/2026/10/506904.html"
tier: "C"
tier_name: "Weak signal"
tier_reason: "QbitAI news write-up that relays DeepSeek's own Super-agent architecture post and prototype repository, with promotional framing and no independent testing; the primary DeepSeek text should be read instead."
summary_basis: "full-text"
original_minutes: "5"
tags: ["ai technology", "agent architecture", "open source"]
---

# 刚刚，DeepSeek重磅开源：Agent底座架构！/ DeepSeek open-sources its Super-agent architecture

Tier C (Weak signal): QbitAI news write-up that relays DeepSeek's own Super-agent architecture post and prototype repository, with promotional framing and no independent testing; the primary DeepSeek text should be read instead.

Summary based on the full text.

## In short

A QbitAI article (31 October 2026) on a DeepSeek engineering post, "Super-agent: towards more autonomous agent systems", released with a prototype repository (deepseek-ai/dsh-superagent-prototype). The design keeps one agent identity across chat groups, GitHub and email while managing many sessions underneath. A Router receives messages, assigns tasks and holds global context; Workers execute tasks in parallel and subscribe directly to follow-up events. In DeepSeek's small simulation the design finished tasks faster and with a tighter tail than a single-session baseline.

## Key numbers

- 12.4 / 15.7 min: median and 95th-percentile task completion time for Super-agent in DeepSeek's simulated multi-task test
- 29.5 min: 95th-percentile completion time for the single-session-plus-subagent baseline under a neutral prompt
- about 300 messages, 20 work requests: size of the simulated collaboration environment; each configuration run 5 times
- 3 to 5: active sessions one person can usually follow at once, per DeepSeek's observation
- about 30% / about 37%: increase in Router decisions and in cache-read tokens when the event pass-through mechanism is removed

## What it reports

- Motivation: per DeepSeek, one person can follow only 3 to 5 active sessions, so human attention limits multi-task agent work as single-task ability improves.
- Four target properties: continuous (context across sessions and platforms), efficient (parallel tasks without blocking), proactive (watches for changes and acts when needed), persistent (long-running with automatic recovery).
- Subscription and pass-through: a Worker that takes a GitHub PR subscribes to later comments, which reach it without going through the Router.
- Speculative execution: when a bug is discussed in a group, the system tries to reproduce and fix it in the background and only reports if it finds a simple fix; the same pattern for meeting invites (check conflicts, draft a reply for confirmation).
- Two-stage event filter: program rules drop irrelevant messages, then a model decides whether the rest needs action, to save tokens.
- Asynchronous ReAct loop so long tasks do not block new instructions; implemented in DeepSeek Harness but not on by default.
- Compute, state and execution environment are separated: agent logic in the cloud, session state stored independently, the local machine keeps the work environment. A separate Watchdog agent detects crashes, loops and drift.
- Test: about 300 messages and 20 work requests, 5 runs per configuration. Super-agent median 12.4 min, P95 15.7 min; baseline P95 29.5 min. Removing pass-through raised Router decisions by about 30% and cache-read tokens by about 37%.

## How they got it

The numbers are DeepSeek's own preliminary simulation, relayed by QbitAI. The article gives no baseline median, no success or quality measure, no model details and no cost per task.

## Limits

- Second-hand: a news relay of DeepSeek's post, with promotional framing ("AI finds bugs and submits PRs by itself").
- One synthetic environment, five runs per configuration; no real-workload data.
- Time only; correctness of completed tasks and false actions are not reported.
- The library summary cites about 10.4k GitHub stars on 2 November; that figure is not in the article and was not checked here.
- The repository is described as a prototype; the library title's "open-sources its Super-agent architecture" overstates it slightly.

## Why it matters for companies

The design addresses a real limit in agent use: people lose track beyond a handful of parallel sessions. Router-Worker separation, event subscription, cheap rule-based filtering before model calls, and a watchdog are reusable patterns for anyone running agents across chat, code and email. Proactive background work also raises control questions the article does not cover: which actions run without approval, what the token bill looks like, and how silent failed attempts are audited.

## Key passages in translation

> 根据他们的观察，一个人通常只能同时跟进3到5个活跃Session。

According to their observation, one person can usually follow only 3 to 5 active sessions at the same time.

> 先通过程序规则筛除无关消息，再由大模型判断剩余事件是否需要采取行动，减少不必要的Token消耗。

Program rules first filter out irrelevant messages, then the large model decides whether the remaining events need action, reducing unnecessary token use.

> 该机制已在DeepSeek Harness中实现，但尚未默认开启。

The mechanism is already implemented in DeepSeek Harness but is not yet enabled by default.

Original: https://www.qbitai.com/2026/10/506904.html