keli-wen/agy-staff

Hire Google's Antigravity CLI (agy) as a fast Gemini staffer for Claude Code and OpenAI Codex.

What it solves

It allows senior AI coding agents (like Claude Code or OpenAI Codex) to delegate time-consuming or repetitive tasks to a faster, secondary worker agent powered by Google's Antigravity CLI (agy) and Gemini 3.8 Flash. This reduces the load on expensive or slower primary models, providing a "fast lane" for research, reviews, and scoped implementations while the main agent orchestrates the overall project.

How it works

agy-staff acts as a plugin for various agent harnesses. It introduces five specialized personas—staffer, researcher, reviewer, implementer, and ask—that the primary agent can invoke via specific commands (e.g., /agy:persona in Claude Code).

When a task is delegated, the worker agent runs as a background job with its own execution deadline. The primary agent can then wait for the final result, observe progress through snapshots of tool activity, or cancel/restart the job. The system is designed to be portable, using bash and skills without hard dependencies on a specific runtime library.

Who it’s for

Developers using AI coding agents such as Claude Code, OpenAI Codex, Pi, or OpenCode who want to increase productivity by offloading sub-tasks to a faster model family for better speed and robustness.

Highlights

  • Specialized Personas: Includes dedicated roles for general tasks (staffer), deep-dives (researcher), code/plan audits (reviewer), and focused coding (implementer).
  • Multi-Harness Support: Compatible with Claude Code, OpenAI Codex, Pi, and OpenCode.
  • Background Job Management: Supports asynchronous execution with the ability to track status, observe activity, and set timeouts.
  • Cross-Model Coverage: Leverages Gemini 3.8 Flash to provide a second opinion and robustness that a single model family cannot provide.

Related

  • Project
    google-antigravity/antigravity-sdk-python

    A Python SDK for building Gemini‑powered or locally‑run AI agents. It provides high‑level `Agent`/`Conversation` APIs, streaming responses, multimodal inputs, custom tool registration, safety policies, and support for both Google Vertex AI and offline Gemma‑4 models via LiteRT‑LM.

  • Dispatch
    Gemini 3.5 Flash release notes / what's new

    Google DeepMind has released Gemini 3.5 Flash, a model optimized for agentic workflows and coding that delivers frontier-level intelligence with 4x faster output speeds than other frontier models.

  • Project
    badrisnarayanan/antigravity-claude-proxy

    A proxy server that exposes an Anthropic-compatible API backed by Antigravity's Cloud Code, enabling the use of Claude and Gemini models with the Claude Code CLI.

  • Dispatch
    The Transition from Gemini CLI to Antigravity: A Case Study in Developer Friction

    Google is sunsetting Gemini CLI in favor of the Antigravity platform, sparking a wave of criticism from developers over product stability and open-source regressions.

  • Project
    teng-lin/notebooklm-py

    An unofficial Python client/CLI for Google’s Gemini Notebook (formerly Notebook LM). It lets you program‑matically create notebooks, add many source types, ask Gemini‑powered questions, generate audio/video/slides/quiz/flashcards/mind‑maps, and batch‑download the results. Designed for research automation, LLM‑agent integration (Claude Code, Codex, etc.), and head‑less scripting via a Python API, MCP server, or REST server. Uses undocumented Google APIs, so it may break and is rate‑limited.