agy-delegate
NewDelegate token-heavy but low-difficulty work to a low-cost LLM via the Antigravity CLI (`agy`), so the bulk reading happens outside the main context and only a digest comes back. Use whenever a task meets all three conditions — (a) the expected reading volume is large, (b) each step needs no deep reasoning, (c) only the final summary or list matters. Typical cases — full-codebase inventories (API endpoints, env vars, TODOs), digesting large logs or data files, whole-repo architectural summaries, mechanical compliance sweeps across many files, and bulk web research (surveying many pages and returning a consolidated digest). For locating code or answering "where is X", use the harness's native search or exploration tools (e.g. Explore in Claude Code) instead.
Overview
Delegate to agy
Run the Antigravity CLI (agy) directly via Bash to offload token-heavy, low-reasoning work to a low-cost LLM. You construct the prompt, agy does the bulk reading (files or web), and only its final digest enters your context.
Pick the model
Run agy models first, then choose the best Gemini model: newest version number at the highest reasoning tier (e.g. prefer a newer-generation Flash at "(High)" over older Pro models). Stick to Gemini entries — non-Gemini models in the list defeat the cost-saving purpose.
Build the command
agy -p "<prompt>" \
--model "<best Gemini model>" \
--dangerously-skip-permissions \
--print-timeout 15m- •
-pruns non-interactively and prints the response. - •
--dangerously-skip-permissionsis required in print mode, otherwise tool permission prompts hang the run; for analysis-only tasks this is safe, and you can add--sandboxfor extra restriction. - •
--add-dir <path>(repeatable) brings extra directories into the workspace when the task spans multiple locations. - •Raise
--print-timeoutfor large jobs (default 5m). - •
agycan search the web on its own — no extra flag needed; just ask for it in the prompt.
Write a self-contained prompt
agy sees only what you pass it. Include in the prompt:
- The working directory or target paths, and what to read or search.
- The exact question or extraction task.
- The expected output format (e.g. "a markdown table with columns X, Y" or "a flat list, one item per line") — you consume this output directly, so make it easy to parse.
- An explicit instruction to be exhaustive when the task is an inventory or sweep — low-cost models stop early unless told not to.
Independent jobs run in parallel
When delegating several unrelated sweeps, launch each agy command as a background Bash task instead of running them sequentially.
Treat results accordingly
The output comes from a low-cost model: reliable for enumeration and digestion, weaker on judgment. Spot-check any claim you are about to act on (e.g. verify one or two listed locations exist) before building further work on top of it.
Install & Usage
mkdir -p .claude/skillsAdd the configuration to .claude/skills/agy-delegate.md
/agy-delegateSecurity Audits
Frequently Asked Questions
What is agy-delegate?
Delegate token-heavy but low-difficulty work to a low-cost LLM via the Antigravity CLI (`agy`), so the bulk reading happens outside the main context and only a digest comes back. Use whenever a task meets all three conditions — (a) the expected reading volume is large, (b) each step needs no deep reasoning, (c) only the final summary or list matters. Typical cases — full-codebase inventories (API endpoints, env vars, TODOs), digesting large logs or data files, whole-repo architectural summaries, mechanical compliance sweeps across many files, and bulk web research (surveying many pages and returning a consolidated digest). For locating code or answering "where is X", use the harness's native search or exploration tools (e.g. Explore in Claude Code) instead.
How to install agy-delegate?
To install agy-delegate: create the skills directory (mkdir -p .claude/skills), then add the config to .claude/skills/agy-delegate.md. Finally, /agy-delegate in Claude Code.
What is agy-delegate best for?
agy-delegate is a other categorized under General. It is designed for: api. Created by kime541200.