Get Out of the Model's Way — Kevin Hou, Google Antigravity

Tech / product concepts & features (from the talk)

Core idea: “Do not interfere with the model” (antigravity philosophy)

• LLMs should act as the “star” if the product is built correctly around them.
• The product should avoid constraining or fighting the model’s work.

Antigravity (Google) as an agent-based coding product

• Designed for technical and non-technical users.
• Launched November 2025, with ongoing development supported by developers.

Antigravity architecture: agent manager vs IDE

• Agent-oriented from the start.
• The Antigravity IDE introduced an agent manager concept to coordinate many agents.
• Antigravity 2.0 split into two separate applications:
  • The IDE
  • A standalone agent manager (for orchestration/coordination)

“2.0 in action” capabilities

• Subagents
• New models
• Work trees
• Scheduled tasks
• Voice mode

Product roadmap driven by “scaling through intelligence”

• As the underlying model improves, the product should automatically get better.
• Advanced model capabilities should become visible in the UI (not hidden or limited).
• The speaker ties this to earlier tooling evolution:
  • 2022: deterministic UI features (autocomplete/chat sidebars) using embeddings, rule files, AST parsing
  • 2024: agents introduce new primitives (e.g., MCPs, special tools, permission systems)
  • 2025: agent manager enables managing many agents in parallel; primitives like skills, hooks, artifacts, and others

“Battle scars” / lessons learned

Examples where user trust was initially threatened—but later improved as models got better:

1. Terminal access

   • Concern about destructive actions.
   • Permission systems + smarter models enabled secure and faster development.

2. Removing a chat sidebar

   • Users disliked losing familiar UI.
   • Adoption increased as the agent paradigm became clearer and more capable.

2026: “Agent Team Era” — new primitives

The talk frames future agent systems as agent orchestration (agent teams, swarms, software factories). Three main primitives were emphasized:

1) Team orchestration via “teamwork” agent commands

• Public preview announced: agent commands in Antigravity.
• Usage: type /teamwork to activate a group of agents to complete a task.

Workflow

• User defines a task (more specific tends to work better; the system may request clarifications).
• A lead agent coordinates a team of specialized roles (frontend, backend, infra, testers, designers, etc.).
• Subagents are created dynamically and can run independently and in parallel.
• Subagents may use different models than the main agent (decision made by the lead agent).

2) Sidecars (plugin protocol / long-running utilities)

• New plugin protocol: sidecars as a universal primitive.
• A sidecar is a long-running process that lets a model “listen” to the outside world and set triggers/events.

Trigger/event examples

• SMS
• webhooks
• cron jobs
• GitHub PR connections
• (and similar external signals)

The speaker notes Antigravity already uses sidecars for scheduled/time-based tasks and says a specification will be released for developers to build their own.

3) Generative UI

• The speaker argues hand-crafted interfaces are becoming obsolete.
• Antigravity can render UI directly inside chat, generated on demand rather than template-based HTML.

Example UI interactions

• Task status with Kanban boards
• Timelines (debug-tool-like)
• Interactive charts: bar charts, graphs, tables

Performance claim

• Mentions Gemini Flash in Antigravity at ~900 tokens/second, enabling near-instant UI generation.

Demos / implementation examples (analysis + scale)

Internal projects: agent team generation

• Internal projects demonstrating agent-team generation included:
  • A photo editor that edits raw photos in-browser
  • A messaging app with familiar UI
• Claim: each project required hundreds of subagents and took about half a day.

Hard test: build an OS kernel

• Built from scratch and ran DOOM (demo referenced from Google I/O).
• Scale/stats:
  • 93 subagents
  • 15,000 requests in 12 hours
  • 2 billion tokens
  • < $1,000 cost

Internal research automation example (benchmarking workflow)

• Target workflow: compare multiple deployment options, analyze deltas, and explain causes.
• Automation described:
  • Agents automate ~90% of the process via natural-language requests for desired evaluation outputs.
  • Agents produce a delta, then launch a specialized research agent proposing 100 hypotheses.
  • Subagents parallelize each hypothesis, then merge results into a single explanation/report.
• Includes generating an interactive generative interface to view/filter/structure results.
• Emphasis on reduced need for manual “gluing”:
  • fewer custom async agent pools
  • fewer judge/config pipeline tasks

Relationship to Gemini model capabilities

• The speaker claims Antigravity product work and model work reinforce each other:
  • Gemini learned team-management from the Flywheel research line and Antigravity usage feedback.
• Gemini 3.5 Flash (released April) is cited as enabling capabilities:
  • Faster/cheaper shifting the Pareto tradeoff between intelligence and task speed/cost
  • “Flash is great at managing teams”

Reviews / guides / tutorials

• Mini-guide for trying the feature:
  • Use /teamwork in Antigravity to enter a mode that activates a team of agents.
  • Provide a task description; the system may request extra info.
  • The lead agent coordinates subagents.

Main speakers / sources

• Kevin Hou (shown as “Kevin Howe” in subtitles) — Google Antigravity lead and speaker.
• Mentions/affiliations referenced:
  • Google DeepMind
  • Gemini / Gemini 3.5 Flash
• Additional referenced colleague/demonstrator: Varun (who demonstrated the kernel/OS work at Google I/O).