AI NewsDev toolsAnnouncement

agy-fast cuts Google Antigravity turns by 34 percent across 55 sessions

An MCP plugin for Google Antigravity batches tool calls and reports 34 percent fewer LLM turns on real sessions.

AI News

Editorial2 min read

LinkedInX
agy-fast README header on GitHub

Image: GitHub

Why it mattersAn Antigravity team can measure its own saving before it adopts the plugin. The benchmark script runs on the laptop and leaves nothing.

Many Antigravity sessions spend most of their time waiting for one tool call at a time. INK-Q has released agy-fast. It is an MCP plugin that batches those calls on the laptop. The repository posts measured savings next to the raw sessions.

What the plugin does

The project runs beside Google Antigravity on the user's own machine. It consolidates many tool calls into single calls. One call can carry up to 20 searches or 50 file edits, per the README. One call can also carry 12 build or test commands, or 40 file symbol outlines. Antigravity's own interface and diffs stay unchanged.

INK-Q lists three Antigravity editions: the CLI, Desktop 2.0 and the IDE. It runs on macOS, Linux and Windows. The licence is MIT. INK-Q reports 223 stars on the repository.

The numbers, with their method

INK-Q publishes the method in docs/ANALYSIS_REPORT.md. The report draws on 55 Antigravity sessions. The sessions hold 8,446 LLM turns and 17,302 tool calls, according to the README.

Across those sessions, the project reports a 33.7 percent cut in LLM turns per 10 actions. It reports a 31.5 percent cut in billable token quota. And it reports a 39.7 percent cut in tool-return upload traffic.

INK-Q breaks the figures out further. A long-context session above 150,000 tokens saves 38.2 percent on turns. A subagent run saves 58.8 percent on turns. Fresh input with no cache hit saves 53.7 percent on upload traffic. Every figure is INK-Q's own.

How a team can check the number

The repository carries a benchmark.mjs script. The README describes it as a local tool. It reads an Antigravity session transcript. It then compares a before and after run. The README says nothing leaves the machine.

A team that doubts the 34 percent number can run the script on last week's sessions. The team can see its own figure before it adopts the plugin. The project needs Node.js version 20 or later, per the README.

Two other numbers in the report are worth naming. The turn cut and the token cut are not the same figure. Fewer round trips save turns, and smaller payloads save tokens. INK-Q reports the two together so a team can see which one matters more on its own workload.

A team on Antigravity Desktop may notice bigger savings than a team on the CLI. Two kinds of session gain the most from batching. A session that opens many files is one. A session that fans work out to a subagent is the other.

The wider context is Google's own work on Antigravity cost. In October Google reported that the Gemini 4 Hydrogen orchestrator cut Antigravity quota by 24 percent. agy-fast's figure sits on top of that. It is measured against an Antigravity session, not against a different orchestrator.

A measured number a reader can reproduce is strong evidence. A vendor quoting its own figure with no method is weak. The method is public here, and the script is in the repository.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project