Antigravity's Agentic Leap: OS Kernels & Doom

Kevin Hou of Google DeepMind introduced the core philosophy of agentic product design, likening it to passing the ball to soccer star Lionel Messi in a critical moment, an analogy for enabling large language models (LLMs) to perform at their best.

This approach dictates that LLMs should be viewed as star players, not merely role players, and the product's design should facilitate their capabilities without interference, allowing the models to operate with maximum autonomy.

Antigravity, initially launched in November 2025, aimed to accelerate developers by providing an integrated tool for both technical and non-technical users.

Version 2.0 of Antigravity marked a significant architectural shift by decoupling the integrated development environment (IDE) from the agent manager, allowing the latter to function as a standalone application and introducing features like sub-agents, work trees, scheduled tasks, and a voice mode.

This means the cutting-edge capabilities of the model should be immediately reflected and made visible within the user experience, ensuring that users benefit from every model enhancement.

The evolution of developer tools illustrates this principle, transitioning from deterministic autocomplete in 2022, to agentic workflows in 2024, and finally to parallel agent management in 2025.

Initially, there were significant concerns about granting AI models direct access to terminals, particularly the risk of catastrophic incidents such as accidental codebase deletion.

However, through strategic investment in safety primitives, including robust permission systems, users were able to leverage AI to build and deploy software faster and more safely, ultimately overcoming initial fears as models matured and demonstrated improved autonomous decision-making in terminal execution.

The Antigravity team initially faced user backlash when they removed the familiar chat sidebar in their Windsurf product, replacing it solely with an agent interface.

Despite initial user resistance, this shift was a deliberate move to align with the evolving industry paradigm, which increasingly favors agentic research and execution, ultimately establishing agentic products as the new norm.

By reframing the IDE as a specialized tool for deep abstraction rather than a central hub, Antigravity 2.0 treats agent orchestration, swarms, and software factories as the future of development, as the team noted that "the IDE is to the agent manager what the debugger was to the IDE."

The IDE is now re-envisioned as a specialized tool, akin to a debugger, useful only when developers need to delve into deeper abstraction layers, while the future of development is anticipated to involve advanced agent orchestration, team collaboration, swarms, and software factories.

Through this interaction, Gemini developed the ability to deconstruct complex tasks and effectively manage teams of agents, gaining valuable insights into multi-agent systems.

These improvements were integrated into Gemini 3.5 Flash, launched in April, which now offers enhanced speed, cost-efficiency, and demonstrates leadership in managing multi-agent systems.

Users can initiate agent teams within Antigravity using the simple '/teamwork' slash command, which then activates a lead agent to manage the collaborative effort.

This lead agent dynamically assembles a team of arbitrary size, assigning specialized roles such as frontend engineers, backend engineers, infrastructure specialists, QA, and design, enabling a comprehensive approach to complex tasks.

Each sub-agent is dynamically generated, operates autonomously, and can even utilize a different underlying AI model than the main agent, while the system employs generative UI to provide real-time updates through interfaces like Kanban boards or timeline views.

This ambitious project utilized Gemini 3.5 Flash and involved 93 sub-agents working over 12 hours, executing 15,000 requests and consuming 2 billion tokens.

The entire endeavor cost under $1,000, showcasing the affordability and power of using AI models and sub-agents for complex engineering challenges, thereby proving the scalability and cost-efficiency of this approach.

Traditional side-by-side evaluation workflows typically involve extensive manual effort, often described as 'Jupyter notebook elbow grease,' requiring researchers to painstakingly analyze and compare data.

Antigravity agents automate approximately 90% of this workflow, allowing researchers to use natural language to query the agent about evaluations, which then analyzes Google's massive codebase to calculate performance deltas.

The system further enhances efficiency by spawning a specialist agent to generate over 100 hypotheses for observed findings, with parallel sub-agents investigating each hypothesis, and a generative UI enabling users to filter and interact with the resulting data intuitively.

Antigravity has replaced manual 'hand-engineering' tasks, such as building async agent pools, setting up judges, and taping together data pipelines, with a set of standardized agent primitives.

These primitives are grounded in a fully dynamic sub-agent graph, utilizing a generative UI that tailors findings to individual user learning styles, and allows systems to be reconfigured or regenerated on-the-fly by simply loading a skills file.

In Antigravity, no two sub-agents are identical; the main agent dynamically orchestrates and manages them independently, configuring, prompting, and seeding each sub-agent on the fly.

This dynamic capability allows for parallel operations across diverse environments, including sandboxes and remote systems, leading to increasingly specialized and collaborative agent teams as the underlying AI models continue to improve.

Events like SMS, web hooks, cron jobs, and GitHub PRs trigger actions in Google Antigravity using these long-lived utility processes, which the team defined by stating that "a sidecar process is essentially it is a sidecar process."

These processes can handle diverse external inputs such as SMS messages, web hooks, cron jobs, and GitHub Pull Requests, with Google Antigravity already using this primitive for scheduled, time-based tasks; the full protocol specification is slated for public release later this summer.

Kevin Hou hypothesizes that the era of human-written specialized user interfaces (UIs) is effectively over, advocating for a shift towards dynamically generated interfaces.

Antigravity's Gemini Flash, achieving nearly 900 tokens per second—10 times faster than many competitors—can render UIs inline without relying on static templates or HTML files, enabling highly responsive and adaptive interactions.

This technology supports a wide range of interactive elements, including bar charts, graphs, tables, and even complex applications like running Doom, which Hou compares to Steve Jobs's introduction of the iPhone, signifying a move away from 'fixed in plastic' controls.

The fundamental design principle for agentic products, according to Kevin Hou, is to 'stay out of the way and let the model cook,' emphasizing minimal interference with the model's autonomous operations.

Product owners are encouraged to define flexible primitives that can scale effectively as AI models become faster, more capable, and more cost-efficient, moving away from rigid, heavy mechanical infrastructure.

The ultimate goal is to create a product experience that is not fixed or constrained, allowing the model to adapt and evolve dynamically to meet user needs and leverage its full intelligence.