These docs track main. Latest release: v1.0.0.

antigravity-booster
ProjectDesign history

Antigravity GUI

Research on how the Antigravity desktop app stores plans and brain artifacts.

Historical, superseded

This is a research note kept for design history. It does not describe current behavior; the current docs, starting at Concepts, do. Source: docs/research/antigravity-gui.md.

Condensed from web research. Items marked UNVERIFIED have single-source or conflicting evidence.

Platform shape

Antigravity 2.0 (launched 2026-05-19, Google I/O) is a five-surface platform: standalone agent-first desktop app, agy CLI, Python SDK, Managed Agents API (per-run billing), Enterprise Agent Platform. The legacy VS Code-fork IDE (with Agent Manager) remains available. CLI and desktop app share one agent harness.

  • Agent Manager: spawn/monitor/switch/archive agents across workspaces; Inbox tracks running conversations. One agent per workspace recommended.
  • Brain artifacts: ~/.gemini/antigravity/brain/<conversation-uuid>/ — implementation_plan.md, task.md, walkthrough.md, transcripts in .system_generated/logs/. Plain markdown → readable by anything, including the orchestrator (hybrid-pipeline integration point).
  • Dynamic subagents (v2): primary agent spawns parallel subagents with isolated contexts. A2A protocol supports up-to-16-agent configurations. Docs warn to cap subagent depth (recursion exhausts budgets).
  • v2 commands: /goal (run to completion), /grill-me, /agent (async subagent), /schedule (cron), /btw, /export (CLI → desktop), /usage (quota), JSON hooks (pre-tool, post-edit, session-start).

THE QUOTA REALITY (kills the "infinite tokens" premise)

  • Free: ~20 agent requests/day. AI Pro ($20): ~100/day, weekly refresh since Mar 2026 — overrun → lockout up to 7 days. Ultra $100 (≈5× Pro), Ultra Premium $200 (≈20× Pro, "no weekly cap").
  • Per-model quota pools are separate: Gemini Flash, Gemini Pro, Claude, image gen each independent. Exhausting Gemini does not touch Claude. → Design consequence: route across pools; cross-pool concurrency verified locally (see calibration).
  • Overage credits $0.01 each; credit-per-request conversion undocumented ("ghost-drains" complaint).
  • Top community complaint: background agent loops burning a week's quota; agents re-reading files instead of caching; whole-file regeneration.
  • Quota visibility: /usage (interactive); community tool "Antigravity Cockpit" (VS Code ext) reads per-model quota.

→ Booster design consequences: token/request accounting per run; per-pool concurrency + budget ceilings; prefer diffs over whole-file regeneration in charters; idempotent scheduled tasks (double-fire reports exist); never leave loops unattended without budgets (ADLC Principle 8 already requires this).

Known issues

  • Model label in UI may not match model actually called (UNVERIFIED).
  • Sandbox macOS-only. Linux CLI preview-quality.
  • Scheduled tasks can double-fire — design idempotent.
  • Routing errors: prompt sent to wrong agent tab.

Existing orchestration ecosystem (don't reinvent blindly)

  • oh-my-antigravity (omagy): agy plugin — interview/PRD planning skills, checkpointed execution, omagy team parallel coordination with JSON status. Overlap with booster: team mode. Differentiation: booster is gate-shaped (ADLC), quota-pool-aware, cross-model prosecution, deterministic orchestrator (no model-as-scheduler).
  • Antigravity Cockpit: quota dashboards, auto wake-up reset scheduling.
  • antigravity-awesome-skills, AgentKit 2.0 (16 specialist agents — UNVERIFIED if official).

Source URLs

Roadmap: agy 1.2.8

Integration roadmap written against agy CLI 1.2.8 and ADLC 1.11.1.

Linux sandbox

Research that evaluated Bubblewrap and alternatives for Linux gate sandboxing.

On this page

Platform shapeTHE QUOTA REALITY (kills the "infinite tokens" premise)Known issuesExisting orchestration ecosystem (don't reinvent blindly)Source URLs