Twelve dimensions across seven providers. Re-graded as announcements land, with the market's reaction shown alongside.
Can it actually do work? · updated 2026-06-01
Project Mariner / Astra-class agents are real and improving, if less battle-tested than rivals.
Can the system take a goal and complete multi-step work autonomously (browse, write code, operate tools, recover from errors), not just answer questions? Measures shipped agentic behavior, not demos.
Google pivots from conversational chat toward autonomous, multi-hour agentic task execution.
A continuously-running cloud agent that uses existing Gmail/Workspace access to act on a user's behalf.
Antigravity 2.0's agent-first architecture for orchestrating multiple coding agents asynchronously.
A models/apps/harnesses framework: the harness that lets a model take real actions now matters as much as IQ.
AI engineering matures into a vertical; agents, reasoning, and orchestration tooling are the practical frontier.
Google's intent to fold Project Astra and Mariner into Gemini as a 'universal assistant' that sees, acts, and remembers.