Get Out of the Model's Way — Kevin Hou, Google Antigravity
93 subagents, 12 hours, 2 billion tokens and under $1,000: that's what it took Google's Antigravity to build an operating system kernel from scratch that runs Doom.
Kevin Hou, who leads engineering on Antigravity, Google DeepMind's agentic coding product, explains the principle behind it: "Give Messi the ball and get out of the way." Your product should get better every time the model does. Hou traces how coding tools evolved, from autocomplete in 2022 to agents in 2024, agent managers in 2025 and agent teams in 2026. He also shares the lessons from hard calls like removing chat from Windsurf. He then introduces Antigravity 2.0 and the three building blocks of the agent-teams era. The first is dynamic subagents, where Gemini 3.5 Flash leads a team through /teamwork. The second is sidecars, a new plugin protocol that lets agents listen for webhooks, cron jobs and messages. The third is generative UI rendered on the fly. He also shows how DeepMind researchers automated 90% of eval analysis with 100 parallel hypothesis agents.
The text is the source's own description of its publication. The content belongs to AI Engineer.
More to read
Server-Side Code Execution Tools for AI Agents, Compared
OpenRouter 10 h ago
v0.40.0
Ollama 11 h ago
Google froze its open source bug bounty program due to a ‘significant rise’ in AI submissions
TechCrunch AI 14 h ago
Can ‘super intelligence’ and a non-binding safety pact solve AI’s image problem?
TechCrunch AI 14 h ago