Kevin Hou told AI Engineer that Google Antigravity 2.0 separates the agent manager from the IDE, letting models run their own teams.
He described the new primitives as dynamic sub-agents that spin up in parallel, each with its own model and environment, plus generative UI from Gemini 3.5 Flash that can spit out dashboards at nearly 900 tokens per second, and long-lived sidecars that listen for webhooks, SMS and scheduled tasks.
The showcase leaned on a hero demo that built an OS kernel and ran Doom with 93 sub-agents over 12 hours for under $1,000, but without code, benchmarks or failure modes to inspect, which leaves reproducibility unclear.
Sidecars that listen to the outside world and an agent that freely spins up sandboxes and remote executors widen the attack surface, yet Hou offered no permission model, audit trail or patch story beyond the promise that smarter Google DeepMind models make better choices.
Antigravity CLI brings multi-step reasoning, multi-file editing, tool calling and persistent history directly to the terminal.