Super-Agent Architecture Converges: Five Labs, Six Weeks, One Baseline

Super-agent Applications at Major Model Vendors

Author: 怕浪猫

Tags: 面试, GitHub, 算法

One-Sentence Conclusion

Super-agent is not a standard product category yet, but an architecture paradigm that is rapidly converging: a persistent coordination layer + multiple parallel sub-agents + cross-system connections + governance boundaries. Each vendor names it differently (dot / Duet / Muse / Gemini agent / Super-agent / Agent 365), but the skeleton is highly consistent. September–October 2026 was a concentrated burst: five leading vendors acted within six weeks.

I. Timeline: A Collective Turn in Six Weeks

Date	Vendor	Event
2026-09-08	Meta	Launched personal agent Muse (standalone app)
2026-09-29	OpenAI	DevDay 2026 released always-on personal assistant dots, evolved from Codex
2026-10-08	Google Cloud	Gemini at Work 2026 released general-purpose work agent Gemini agent
2026-10-21	Anthropic	Released proactive agent Claude Duet (research preview, based on Opus 5.5)
2026-10-31	DeepSeek	Released 《Super-agent设计探索》 and open-sourced prototype dsh-superagent-prototype

This is not a coincidence, but the result of the same judgement: model capability is already sufficient; the bottleneck has shifted to "human attention" and "conversation management".

II. Overseas Vendors, One by One

Meta · Muse

Form: C-end personal agent app, all platforms.
Data: Q3 earnings call on 2026-10-21 disclosed over 10 million users; about 25% of activated users use it daily; more than a quarter of active shopping users have an average order value > $200. Sensor Tower counts 10.9 million downloads (since 9/8 launch).
Key design: Uses Plaid connectors to link bank accounts for agentic commerce, with Stripe's Link for checkout (Meta claims it is the first AI agent covered by Link purchase protection); on the privacy side there is Muse Secure VM, and a planned Confidential VM promises that even Meta itself cannot see user content.
Monetization: Most users are free; revenue comes from taking a cut of in-platform transactions, and heavy users pay for premium models/more tokens. Zuckerberg said he wants to improve compute cost efficiency by more than 10x within a year.
Controversy: Muse has been criticized for building profiles of users' friends and family—it can recommend restaurants, but also feed the advertising business.

OpenAI · dots

Form: A resident assistant embedded in ChatGPT desktop and mobile, not a standalone app. The name comes from the negative space inside the OpenAI logo.
Technical origin: Evolved from the Codex programming tool. Enterprise product lead Alexander Embiricos explicitly said Codex could already do most of what dots does, but required a lot of manual prompting.
Key design: Comes with its own cloud computer and browser, can run 7×24 in the background; connects to 4000+ apps; "proactive research" periodically reorganizes known user information, history, and authorized resources and gives action suggestions, but only with read-only tools from connected apps—it cannot send messages, modify files, or take over the computer; Space enables interoperability between cloud dots and local Codex content; users can hit "take over" to operate the cloud computer themselves.
Availability: Pro / Business Premium users in eligible markets are gradually granted access; Enterprise requires admin enablement; there are enterprise-level Specialist Dots. Subscription starts at $100/month.
Known weaknesses (self-stated): poor alert timing (Embiricos cites "schedule conflict alerts arriving only 3 minutes before a meeting"); coarse permission granularity (does not accept read-only authorization for cloud drives), and the official acknowledgment that "super, super, super granular" controls are needed.
In talks: discussing payment-pass-through capabilities with several payment processors; partnering with Microsoft to bring dots under Agent 365 security governance.

Google · Gemini agent

Form: Converges QA, knowledge work, media creation, and code writing/execution into the same agent, the same prompt box plus an API.
Four architecture principles: unified form; omnichannel access (Web/iOS/Android/Win/Mac/CLI/Workspace/M365/Slack, plus headless embedding into third parties); cloud-persistent execution (tasks continue for hours to days after closing the laptop); multi-agent orchestration.
Two most notable points:
1. "Coworker agents": take on persistent roles like project manager or financial analyst inside a team, with a dedicated @agents.company.com email address and persistent storage, only able to access explicitly limited data and channels, executing tasks across days and sessions.
2. Model choice: the agent picks a model by default; beyond the Gemini series, Anthropic's Claude is already callable, with Smart Routing and real-time spending caps to control cost.
Connections: MCP connects Workspace, M365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Snowflake, etc.; there is an enterprise tool registry for teams to publish custom tools.
Status: private preview for Workspace enterprise customers; finance and legal editions are in preview, government/healthcare/retail editions are on the way. Pichai disclosed Gemini MAU is over 1 billion, and nearly 90% of Fortune 100 companies use Gemini Enterprise.

Anthropic · Claude Duet (after Cowork)

Form: standalone app (iOS / Android / claude.ai web), based on Claude Opus 5.5.
Key design: inherits memory, skills, and connected apps from the Claude account, no reconfiguration; comes with a built-in VM and browser so it can run even when the device is off; can call over 6,000 connectors from Claude Marketplace (Gmail, Slack, QuickBooks, Stripe, Gusto, HubSpot, Canva, Shopify, DocuSign, etc.).
Measure cadence: a daily morning briefing, then patrols every few hours during the day—officially explicit about not doing continuous patrols, because that would quickly consume plan quota, and patrols count toward shared usage.
Permission tiers: reading, drafting, preparing can be autonomous; before sending messages, making payments, signing documents, or deleting content, the full plan must be shown to the user for approval; buying/selling securities and transfers between accounts are not done even with approval; credit card numbers and government ID numbers are directly rejected by a dedicated classifier.
Status: U.S. Pro/Max waitlists are released in batches; research preview does not cost extra; officially admits it can misread conversations and misjudge urgency, is weak at personal errands (like booking tickets), and does not currently show browser operation steps.

Microsoft · Agent 365 (not an agent, but a control plane)

Ignite 2026's main theme is "from Copilot to autonomous agents": M365 and Azure autonomous Copilot agents reach GA.
Copilot Studio gets agent-to-agent orchestration, MCP connectors (ServiceNow / SAP / Salesforce), governance dashboards, version management, and rollback.
Agent 365 is positioned as the control plane for AI agents—centralized deployment, governance, security policies, and access control. This is the route of productizing the "governance layer" separately.
Licensing: M365 Copilot ~€30/user/month, with additional Copilot Studio capacity units for high concurrency.

DeepSeek · Super-agent

Positioned as an architecture exploration for R&D scenarios, not a consumer product. Router (persistent coordination, only routes without executing) + Worker (each has independent context, in parallel) two-layer structure, event-driven + two-stage filtering (rules → model semantic judgement), asynchronous background, Watchdog not only monitors downtime but also whether tasks have gone off track, context and execution environment decoupled (brain in the cloud, records stored separately, code left on local machine).
The prototype dsh-superagent-prototype is open-sourced, but officially it is a research prototype, currently targeting a single GitHub repository, with optional Feishu integration; some designs in the article are not yet implemented.

III. Domestic (Chinese) Vendors: Three Different Routes

Note: The domestic section lacks official first-party disclosures from each vendor; the following is synthesized from H1 2026 industry reviews, local industrial-policy articles, and selection guides. Data reliability is weaker than the overseas section; please verify against official announcements.

Vendor	Carrier	Route characteristics
Alibaba	Tongyi Qianwen Qwen-Agent 3.0 (super entry) + Alibaba Cloud Bailian	Ecosystem closed loop: connects Taobao, Fliggy, Alipay, AMap, DingTalk "Wukong", pushing Agentic Commerce (AI directly places orders, pays, and tracks logistics); Tongyi MAU exceeds 100 million. Bailian transforms into a "model supermarket", uniformly accessing GLM-5.1, MiniMax M2.7, Kimi K2.6 and over 160 models
ByteDance	Doubao (C-end) + Coze (B-end low-code)	Terminal / system-level: Doubao Phone Assistant 2.0 deeply integrates into the OS, self-developed cross-domain scheduling, edge-first for privacy hardening, launched second-generation Doubao phone in July; Coze 3.0 supports multi-person multi-agent collaboration, Agent Skills, and long-term plans
Tencent	Hunyuan + Yuanqi + WorkBuddy / CodeBuddy	Private domain and office: Hy-Memory long-term memory solves cross-session forgetting; Yuanqi accumulates 70,000+ industry skill components; WeChat/WeCom ecosystem is the main moat
Moonshot	Kimi K3 (2.8 trillion parameters, 1M context) + Kimi Work	Developer / knowledge work: model-API call revenue accounts for over 70%, overseas paying users grew 4x
ZhiPu	GLM-5.2 + AutoClaw / AutoGLM	Native agent: can complete full workflows autonomously without third-party plugins; open-sourced 60+ models under MIT license, self-developed ZClawBench evaluation
Baidu	Qianfan + DuMate	Proposes DAA (daily active agents) metric, Kunlun chips for domestic compute closed loop
Huawei	AgentArts + Pangu	Industrial, mining, energy; full-stack hardware/software autonomy
iFlytek	Xingchen Agent Platform	The only one that directly uses the name "SuperAgent": marketing agent platform, proposes S.U.P.E.R five-dimension capability standard (Symbiotic / Universal / Production / Evolution / Reasoning), emphasizing the shift from single-agent to swarm intelligence

Domestic vs US emphasis differs clearly: the US leans toward enterprise software integration and governance; China leans toward consumer scenarios and super-app ecosystems.

IV. Converged Commonalities (de facto standards)

Cloud VM / sandbox: tasks keep running even when the device is off—Muse, dots, Duet, Gemini agent all do this.
Persistent memory across sessions: no more "forgetting when switching conversations". Tencent Hy-Memory, Kimi hierarchical memory, Duet inheriting account memory are all the same thing.
MCP becomes the de facto standard for tool connections, A2A (agent-to-agent) is just starting, ByteDance Coze already supports it.
Multi-agent orchestration: dynamically creates sub-agents with independent identities, parallel or serial. Google's "coworker agents" given email and persistent storage are the most aggressive approach.
Tiered permissions with highly consistent red lines: reading/drafting/preparing can be autonomous; sending, paying, signing, deleting require manual approval; securities trading and inter-account transfers are a common taboo across vendors (Duet explicitly says it won't do them even with approval).
"Keeping quiet" is a product capability: does not broadcast the process, only speaks when results are ready, a human decision is needed, or the task cannot be done. DeepSeek's wording is most blunt—"occupy as little human attention as possible, provide as much value as possible".
Cost is a hard constraint, not an afterthought: Duet avoids continuous patrols out of fear of burning quota; Google uses Smart Routing + real-time spending caps; Meta wants 10x cost efficiency in a year. Super-agent essentially spends tokens in exchange for human attention.

V. Unresolved Divergences

Divergence point	Status
Boundary of proactivity	dots reminds about schedule conflicts only 3 minutes before a meeting; Duet can misjudge task urgency. "When not to interrupt" remains unsolved
Permission granularity	dots does not accept read-only authorization for cloud drives; the official concedes control granularity is insufficient; users must weigh "useful" against "handing over privacy"
Can agents socialize with each other	Meta wants Muse network (your assistant directly connects with merchant assistants); OpenAI's Embiricos flatly said "I personally have no intention of meeting with someone else's AI agents", and still hires a human assistant because partners require a human counterpart
Business model	Commission (Meta) vs subscription (OpenAI/Anthropic) vs enterprise licensing (Microsoft/Google)
Merchant-side acceptance	Whether websites, payment networks, and merchants are willing to host AI agents—the whole industry has not yet cracked it

Domestically there are two harder bottlenecks: the stability of autonomous decision-making in complex non-standard scenarios (unattended full-process landing cases are under 15%), and legacy ERP / MES / government systems lack standard open interfaces; customization costs shut out small and mid-sized enterprises.

VI. Your Judgement

If you are building R&D tools: DeepSeek's Router/Worker layering is worth borrowing directly, especially "Router is extremely light, Worker context is isolated, and conflicts are detected by subscribing to the same Thread"—it avoids the problem of heavy inference overhead and high-frequency coordination when a main agent leads sub-agents.
If you are building enterprise-side products: productizing the governance layer separately (Microsoft Agent 365 route) builds a stronger moat than racing on the agent itself. Google building identity, email, persistent storage, and spending caps as built-in components is the same idea.
In the short term, do not bet on a single route: proactivity, permission granularity, and agent-to-agent interaction are all still being tried across the industry, and even the leading vendors are changing their defaults.