Introduction
Why agb exists and how the docs are organized
antigravity-booster (agb) makes Google Antigravity (the agy CLI and the desktop app) work for large parallel build-outs. It applies the Agentic Development Lifecycle (ADLC) on top of Antigravity: deterministic orchestration, scheduling that knows about quota pools, cross-model prosecution, and validation built from gates, covering everything from the plan to the merge.
Why
Antigravity gives you a generous model quota (not an infinite one) split across independent per-model pools. Running work in parallel by hand is slow, and unsupervised agents burn through quota fast. agb turns that quota into a disciplined fleet:
- Code handles control flow, models handle judgment. A Node scheduler dispatches
agyworkers, and no model ever decides the order work runs in. - Quota pools multiply concurrency. The Gemini Flash, Gemini Pro and Claude pools throttle and meter independently. The scheduler holds a semaphore per pool and routes tiers across them. See Pools and tiers.
- Cross-model prosecution. A model from a different family than the builder prosecutes each diff under a refute charter with a JSON verdict contract, and
adlc hollow-testmutation evidence backs it up. Critical and high findings block the merge. See Prosecution. - Deterministic gates. Build and test commands gate every ticket in its own worktree, then again after the merge on the base branch. A failed post-merge gate reverts the merge. See Gates and sandboxing.
- Two-strike regeneration, gated by
adlc flail-detector. A failed first strike is regenerated from scratch with the failure diagnostic; ifadlc flail-detectordetects a repeated failure pattern, the second strike is halted before burning pool quota. - Frozen rails. Paths a ticket declares as
railscannot be edited while it is in flight. This is enforced in-session by the policy guard and at merge byadlc rails-guard. See Rail enforcement.
Planning stays in Antigravity. agb plan compiles the brain artifact (or a Markdown spec) Antigravity produces into a gated plan.json, and agb run executes it. See Plan boundary.
Measured impact
Across 100 paired runs per workload (npm run eval, calibrated against live CLI probes), agb outperforms both a default single-session agy workflow and a well-tuned agy + best practices setup (/plan + /teamwork + shared worktrees + adversarial-reviewer subagents):
| Workload | Default agy | agy + best practices | With agb | Gain vs. best practices |
|---|---|---|---|---|
| 16-ticket feature build | 12m 12s (1.31 / min) | 8m 59s (1.78 / min) | 3m 49s (4.20 / min) | 2.36x faster (saves ~5m 10s) |
| 8-agent parallel burst | 68.5s (25.8% timeouts) | 57.2s (4.2% timeouts) | 30.8s (0% timeouts) | 1.85x faster, 0% drops |
| Pre-merge bug detection | 28.4% caught (71.6% escape) | 74.6% caught (25.4% escape) | 94.6% caught (5.4% escape) | 4.7x fewer escaped bugs |
| Protected-file & scope safety | 31.0% blocked (207 breaches) | 74.0% blocked (78 breaches) | 100% blocked (0 breaches) | Zero rail breaches |
See Benchmarks & Statistical Evals for the full methodology, hypothesis tests, Pareto frontier charts, and 80/20 root-cause breakdowns.
How these docs are organized
Getting Started
Install the plugin, bootstrap, run your first plan, and verify your environment.
Concepts
The four layers, quota pools, gates, prosecution and rail enforcement.
Benchmarks & Evals
How agb compares to default and best-practices Antigravity setups.
Guides
Task-oriented walkthroughs: writing plans, sweeps, review, the sidecar dashboard.
Reference
Every command, flag, schema, environment variable and exit code.
Internals
How agb is built: run lifecycle, integration journal, plugin and migration.
Project
Stability policy, contributing, releasing and design history.
New here?
Start with Installation, then Quickstart. If you already have the npm package installed, read Upgrading from npm.