This week's strongest zero-human company signals came from three geographies at once: Paris-backed collaborative AI infrastructure, Google's attempt to formalize the agentic web stack, and Alibaba's push for longer-running agent execution in China. The pattern is the same across all three. The market is moving from isolated assistants to shared systems that can coordinate, transact, and stay on task.
1. Investments: Capital Is Rewarding Shared Agent Systems
On May 18, 2026, Paris-founded Dust announced a $40 million Series B with Abstract, Sequoia, Snowflake Ventures, and Datadog. Dust says it now serves more than 3,000 organizations globally, with more than 300,000 deployed agents and a thesis it calls "multiplayer AI."
That phrase matters. The investment is not a bet on another personal copilot. It is a bet that companies will need a shared layer where humans and agents can coordinate around the same context, files, tools, approvals, and goals. That extends the investment signal we covered in Polsia's $30 million raise from the one-person company outward into the multi-agent company operating layer.
2. Frameworks: Google Is Trying to Standardize the Agentic Web
Google I/O landed a more important standards story than most people noticed. On May 19, 2026, Google said its new WebMCP proposal is designed to let sites expose structured tools to browser agents with explicit schemas, discovery, and page-state awareness. In practical terms, that is Google trying to reduce the brittleness of click-and-guess browser automation.
This sits in the same family of work as our earlier coverage of A2A and Microsoft's governance toolkit. A2A helps agents coordinate with each other. Governance frameworks help control them. WebMCP starts defining how websites should talk back to agents in a structured way.
3. Tooling: The Workspace and Terminal Are Collapsing into One Agent Surface
Google paired the standards move with a tooling move. At I/O, it introduced Antigravity 2.0 and Antigravity CLI, positioning them as a common harness for multi-agent orchestration, subagents, scheduled background work, and agent deployment across developer surfaces.
A day earlier, on May 13, 2026, Notion introduced its Developer Platform, including hosted Workers, an External Agent API, and a CLI that makes outside agents first-class participants inside a shared workspace.
Put together, these launches matter because zero-human companies do not just need better models. They need durable control planes where agents can run in the background, enter the same workspaces as humans, and act across systems without every team rebuilding the harness from scratch. That sharpens the thesis from our workspace agents piece: shared operational context is becoming a product category.
4. AI Capabilities: Long-Horizon Execution Is Becoming a Real Differentiator
On May 21, 2026, Alibaba published Qwen3.7-Max as a model built for the agent era: coding, office workflows through MCP, multi-agent orchestration, and sustained execution across hundreds or thousands of steps. Alibaba says the model completed a 35-hour autonomous kernel optimization run with more than 1,000 tool calls.
Google made a different but related claim on May 19, 2026, saying Gemini 3.5 Flash is aimed at high-speed, real-world agentic workflows. The important pattern is not any single benchmark. It is that major labs are now competing on whether a model can keep operating over time, across tools, and across changing context. That is the same shift we highlighted in our GPT-5.5 field notes.
5. The Global Pattern
The regional spread is worth noting. Dust gives Europe a stronger claim on the organizational AI layer. Google is trying to define the default agent interfaces for the web and developer workflow stack. Alibaba is pushing a China-centered full-stack answer that spans models, chips, cloud, and agent scaffolds.
That is what category formation looks like. The question is no longer whether agents can do interesting demos. The question is which regions and platforms will own the standards, operating surfaces, and execution layers that zero-human companies run on.
6. What Changed Since Our Last Coverage
Our April 29 briefing argued that the stack was widening from models into infrastructure, standards, and creative tooling. This week extends that thesis with more precision.
Capital is now rewarding shared human-agent systems. Browser standards are being rewritten for agent actuation. Workspaces are being redesigned to host external agents natively. And frontier model launches are emphasizing autonomous execution time, not just answer quality.
That is a more mature stack than we had a month ago.
Related: See our prior research on the April 29 briefing, Polsia, workspace agents, A2A, and GPT-5.5.