ADLC orchestration for Google Antigravity

Mission control for your agent fleet.

agb runs large parallel build-outs on Antigravity without losing control: a deterministic scheduler, frozen rails, sandboxed gates and cross-model prosecution between every agent and your main branch.

$agy plugin install https://github.com/voodootikigod/antigravity-booster.git

Read the docs Quickstart Benchmarks & evals

Flight sequence · one ticket, start to merge

01Planagb plan
Compile a spec into a ticket DAG, then gate it for overlap and executability.
02Dispatchquota pools
Schedule tickets deterministically across model pools, inside the caps.
03Buildagb/<ticket>
Each ticket builds in its own worktree, with frozen rails enforced live.
04Gatesandboxed
Tests and checks run in a network-denied sandbox. Failure stops the ticket.
05Prosecutecross-model
A different model family tries to break the change before it can merge.
06Mergerebase-first
Survivors land one at a time, gated again in an integration worktree.

Benchmarks · 100 paired runs per workload

Faster builds, fewer escaped bugs—even against a well-tuned setup.

You can get a lot out of Antigravity on its own by combining /plan, /teamwork, shared worktrees, and an adversarial-reviewer subagent. We benchmarked agb against both a default single-session workflow and that hand-tuned best-practices setup to measure what deterministic scheduling, multi-pool routing, and mechanical gates actually add.

Workload	Default agy	agy + best practices	With agb	Gain vs. best practices
16-ticket feature build Parallel DAG from spec to main	12m 12s (1.31/min)	8m 59s (1.78/min)	3m 49s (4.20/min)	2.36x faster Saves ~5m 10s per run
8-agent parallel burst Concurrent dispatch under quota limits	68.5s (25.8% timeouts)	57.2s (4.2% timeouts)	30.8s (0% timeouts)	1.85x faster Zero rate-limit drops
Pre-merge bug detection 500 seeded bugs across 5 categories	28.4% caught (71.6% escape)	74.6% caught (25.4% escape)	94.6% caught (5.4% escape)	4.7x fewer escaped bugs -20.0 percentage points
Protected-file & scope safety 300 attempts to touch frozen files or skip tests	31.0% blocked (207 breaches)	74.0% blocked (78 breaches)	100% blocked (0 breaches)	Zero rail breaches Enforced by hooks, not prompts

Speed vs. code quality

Usually, running more agents in parallel means more broken merges. By spreading work across model pools and reviewing every diff with a second model family, agb raises build speed and bug catch rate at the same time.

Default agy

1.31 / min28.4% caught71.6% escape

Single session or unthrottled burst, self-review

Speed

Bugs caught

agy + best practices

1.78 / min74.6% caught25.4% escape

/plan + /teamwork + shared worktrees + adversarial-reviewer

Speed

Bugs caught

With agb

4.20 / min94.6% caught5.4% escape

Multi-pool scheduler + frozen rails + cross-model prosecution

Speed

Bugs caught

Where the gains come from

Over 80% of the time savings and bug catches come from two things prompts alone can't do: routing across separate model quota pools in code, and enforcing file boundaries outside the model.

Why builds finish 72% faster (~8m 17s saved per 16-ticket run)

Spreading work across Gemini, Claude, and GPT quota pools258s saved · 52%

Scheduling in Node.js instead of spending LLM turns on /teamwork151s saved · 30%

Keeping each pool under its rate-limit knee (no 120s timeouts)55s saved · 11%

Serial rebase-first merge lock + halting repeated failures early32s saved · 7%

Why 89% of seeded bugs & rail violations are stopped

Cross-model prosecution + 4-specialist review before merge53% of catches

PreToolUse hook + rails-guard blocking edits to frozen files29% of catches

Mutation testing (hollow-test) + sandboxed test runs18% of catches

Run the benchmark suite yourself:npm run evalnode evals/runner.mjs --json

Full methodology, statistical tests & Pareto charts →Live CLI calibration data →