# Agents Built an OS Kernel That Runs Doom — Kevin Hou, Google DeepMind

## Executive summary

Kevin Hou details the evolution of agentic coding tools, introducing Antigravity 2.0, Google's agent coding product. The core principle is 'scaling with intelligence,' meaning product primitives must evolve as LLMs improve. The future (2026 era) is defined by 'agent teams' or 'swarms,' which are powered by three new primitives: dynamic subagents, sidecars, and generative UI. These tools allow for complex, parallelized tasks, demonstrated by building an entire OS kernel that runs Doom using 93 subagents in 12 hours.

## Key takeaways

- The Agentic Evolution: Coding tools have progressed from deterministic autocomplete (2022) to agent-based systems (2024) and parallel agent managers (2025). The next phase (2026) is focused on agent teams/swarms, which decouple the IDE from the agent manager for greater flexibility.
- The Three Core Primitives: The future of agent teams relies on: 1) Dynamic Subagents (specialized, parallel roles); 2) Sidecars (a new plugin protocol for listening to external triggers like webhooks or cron jobs); and 3) Generative UI (rendering interactive UIs on the fly, bypassing fixed templates).
- Demonstrated Capability: Antigravity successfully built a complete OS kernel from scratch and ran Doom. This feat required 93 subagents over 12 hours, utilized 2 billion tokens, and cost under $1,000.

## Technical details

- Antigravity 2.0 Architecture: The product was updated by decoupling the IDE from the agent manager, allowing the agent manager to function as a standalone 'mission control' application.
- Dynamic Subagents: Subagents are dynamically generated and configured by a main orchestrating agent. They can operate in parallel, take on specialized roles (e.g., front-end, QA), and even select different models from the main agent's model.
- Sidecars Protocol: Sidecars are a new plugin primitive enabling the model to listen to the outside world. This allows agents to set up triggers based on external events such as SMS messages, webhooks, or GitHub PRs.
- Generative UI: This primitive allows the system to render interactive UIs (e.g., Kanban boards, graphs, data filters) on the fly, rather than relying on fixed HTML or templates, improving user interaction and data visualization.
- Model Capabilities: The system leverages Gemini 3.5 Flash, which is noted for being fast, cheaper, and highly capable of leading and managing teams of agents.

## Practical implications

- Product builders must design systems that scale with the underlying model's intelligence, rather than relying on fixed UI/UX patterns.
- The decoupling of core components (like IDE and agent manager) is critical for building robust, modular agentic applications.
- The sidecar protocol offers a powerful mechanism for integrating LLM workflows with existing enterprise systems (e.g., GitHub, CRON jobs).
- Complex workflows, such as automated research analysis (side-by-side evals), can be automated by combining subagents and generative UI, drastically reducing manual 'elbow grease.'

## Topics

Agentic Workflow, LLM Architecture, Product Engineering, DevTools, AI Development, Microservices, Google Antigravity, Gemini 3.5 Flash

Source: https://www.youtube.com/watch?v=buHC7bQE1X4