Super-Agent Architecture Converges: Five Labs, Six Weeks, One Baseline
Any developer building agent workflows now inherits an emerging industry baseline: cloud VMs, persistent memory, MCP, and mandatory human sign-offs on irreversible actions. Ignoring those defaults means shipping agents that die with a closed laptop, forget context, or get blocked by payment and security teams.
The super-agent boom broke in autumn 2026. Starting September 8, Meta launched Muse; OpenAI followed with dots at DevDay (Sept 29), Google with Gemini agent (Oct 8), Anthropic with Claude Duet (Oct 21), and DeepSeek with an open-sourced prototype (Oct 31). The five products share a de facto architecture: a persistent coordinator, independent sub-agents, a cloud sandbox that survives a closed laptop, cross-session memory, and MCP as the tool connector. The differences are mostly naming.
Two things stand out. First, every vendor treats cost as a design constraint, not an afterthought: Duet only checks in a few times a day to spare quota, Google adds smart routing and spending caps, Meta promises a 10x cost-efficiency increase. Second, permission boundaries are nearly identical across the board: reading, drafting, preparing are autonomous; sending, paying, signing, deleting require human approval; securities trading and cross-account transfers are prohibited even when approved. That convergence is a settlement, not a coincidence.
Divergences remain around proactivity timing, permission granularity, and whether agents should talk to other vendors' agents. OpenAI's product head said he personally would never attend a meeting with someone else's AI. Meanwhile Chinese vendors are splitting along product lines: Alibaba's super-app integration, ByteDance's OS-level phone assistant, Tencent's WeChat ecosystem, and more. The U.S. leans enterprise integration and governance; China leans consumer super-apps.
The six-week cluster points to a shared threshold: model capability is no longer the limiting factor, so attention and session management became the bottleneck and every vendor independently shipped the same skeleton.
Google's coworker agents with corporate email addresses and persistent storage turn agents into organizational actors, which forces governance onto the same rails as human employees.
The near-identical red lines on sending, paying, signing, and deleting show the industry settling on risk boundaries before any technical mechanism makes them safe.
OpenAI's dots accepting only read-only tools for proactive research reveals how quickly autonomy collides with accidental writes; the official admission that granular controls are needed is a concession to that reality.
DeepSeek's Router/Worker pattern directly offloads coordination from a heavyweight main agent, addressing the inference cost that bogs down sub-agent orchestration in mainstream designs.
The split between US enterprise-governance focus and Chinese consumer-super-app focus is not just a regional taste; it sets different cost and privacy trade-offs for agents that Western developers will either adopt or work around.