Get Out of the Model's Way — Kevin Hou, Google DeepMind
Source
youtube.com
Author
AI Engineer
Date
Context
Google DeepMind's Antigravity presentation describes sidecars, a plugin protocol that lets coding agents listen for webhooks (automatic notifications from other services), cron jobs (scheduled tasks) and messages, according to the talk description.
That is a concrete extension of the coding-assistant workflow: an arriving event can become an input to AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → work rather than waiting for a person to start it. The talk presents sidecars as one of three building blocks of an agent-teams era, alongside dynamic subagents and generated interfaces. The supported contribution here is the event-listening mechanism, rather than the headline demonstrations' claimed scale or cost.
Key takeaways · AI-distilled
Hou says Antigravity built an OS kernel that runs Doom from scratch using 93 subagents over 12 hours, about 2 billion tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → and under $1,000.
His product principle is to get out of the model's way, so the tool improves every time the model does; he also recounts hard calls such as removing chat from Windsurf.
Antigravity 2.0's building blocks are dynamic multi-agentUsing several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.Full definition → teams led by Gemini 3.5 Flash via /teamwork, sidecars (a plugin protocol letting agents react to webhooks, cron jobs and messages) and generative UI.
He says DeepMind researchers automated about 90% of evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → analysis by running 100 hypothesis agents in parallel.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
videoRethinking Claude Code Best Practices — Lydia Hallie, Anthropic
videoStop Shipping LLM Settings That Only Win on Paper - Rohan Pandey & Balaji Varadarajan, DigitalOcean
videoYour SKILL.md Is 70% Stuff the Model Already Knows - Nikolay Petrov, Intel
videoYour Spec Has Bugs. Catch Them Before Your Agent Codes. - Rajath Mahendrappa Kotyal, AWS
