# agy-staff: delegating work from Claude Code and Codex to a fast Gemini CLI > agy-staff is a plugin set that lets a senior coding agent hand scoped work to Google's Antigravity CLI, agy, running Gemini 3.8 Flash, through five personas. It ships for Claude Code, Codex, Pi and OpenCode, and it comes with its own benchmark harness, whose results say delegation buys speed of opinion but costs wall clock time. **keli-wen/agy-staff** — Hire Google's Antigravity CLI (agy) as a fast Gemini staffer for Claude Code and OpenAI Codex. - Repository: https://github.com/keli-wen/agy-staff - Stars: 788 · Forks: 52 - Language: JavaScript - License: MIT - Published: 2026-10-21 · Updated: 2026-10-21 · Language: en - Canonical page: https://hysenlabs.com/projects/keli-wen-agy-staff ## The pitch is a fast lane beside a slow senior agent agy-staff is a plugin set for coding agents. Its premise is that the model driving your session is fast at reasoning but expensive or slow for anything small, and that a second, cheaper model should handle the small jobs instead. The worker it delegates to is agy, the Antigravity CLI, which ships fast Gemini 3.8 Flash access. The project names five personas. Staffer is the general purpose one, researcher handles search and reading tasks, reviewer covers code and also plans and decisions, implementer does scoped work, and ask is a plain question with no tools. There is also a model facing jobs skill for managing the work queue. Claude Code invokes them with a slash prefix, /agy:, and Codex with a dollar prefix, $agy:. The reasoning in the README is a cost argument rather than a capability argument. Codex is described as slow even in fast mode, Claude Code as quicker but still not fast, and the author's own quota as scarce enough that the senior agent should orchestrate subagents rather than grind through every survey and review itself. The secondary argument is coverage: a second model family looking at the same code catches things the main agent cannot catch by itself. ## The personas run unrestricted by default, and the project says so plainly This is the part of the design that deserves the most attention, and the README puts it under an important notice rather than a footnote. There is no mandatory setup step. The staffer, researcher, reviewer and implementer personas run unrestricted, which means agy can inspect the repository, run commands and edit files. The project defends that choice on practicality grounds, saying the prompts adapt to the current repository state. When implementer starts in a dirty workspace, the companion process tells agy which files already had changes and reminds it not to overwrite or deliver unrelated work. If the task asks for a commit, a push or a pull request, agy can do that delivery, and otherwise it leaves a working tree diff for review. The notice then states the limit plainly, and it is worth quoting the substance: the prompt instructions do not provide permission isolation. Running a second agent with write access to your repository is a trust decision, not a technical one, and the mitigation offered is a setup step with a --restricted flag, described as optional hardening for untrusted input. The ask persona is the exception, since it is tool free and needs no setup. ## Installation is the Antigravity CLI first, then one plugin command The order matters. The Antigravity CLI has to be installed and authenticated first, and the README points at the official Antigravity documentation for that, with agy --version as the check. Node.js is required as well. ```bash curl -fsSL https://antigravity.google/cli/install.sh | bash ``` For Claude Code the plugin is added through its marketplace mechanism. ```bash claude plugin marketplace add keli-wen/agy-staff claude plugin install agy@agy-staff ``` Codex uses a parallel pair of commands with a full URL. ```bash codex plugin marketplace add https://github.com/keli-wen/agy-staff codex plugin add agy@agy-staff ``` After a restart, the first thing to try is the ask persona, since it needs no tools and no setup, so you can confirm the wiring works before letting anything touch your files. Claude Code takes /agy:ask reply with OK and Codex takes $agy:ask reply with OK. Pi and OpenCode have their own instructions in collapsed sections of the README. The OpenCode path installs through the native plugin manager, and the note there is the most useful piece of troubleshooting advice in the whole document: OpenCode caches the complete package spec, so restarting does not refresh an unchanged Git spec, and upgrading means choosing a newer tag or commit and reinstalling with a force flag. The adapter is verified against OpenCode V1, with a stated engine range of greater than or equal to 1.18.34 and below 2, and V2 is explicitly not covered. ## The bench says delegation buys cheap opinions and costs wall clock time agy-staff carries its own evaluation harness, and the results are the most useful part of the project for anyone deciding whether to use it. Bench v1 runs Claude Code and Codex on three benchmarks, SWE-bench Verified, SWE-Atlas QnA and DeepSearchQA, with each task run three ways: alone, with their own subagents, and with agy through the lead persona. The stated purpose of the bench is three questions, and they are worth repeating because they show what a good agent evaluation looks like. When to delegate, meaning which tasks and hosts gain from handing work over and which are better left alone. What to fix in agy-staff, since each run shows where delegation wastes time or tokens and later versions check whether the fixes worked. And how harnesses differ, since the same prompt and the same worker on different hosts trace the gaps back to design choices in each harness. The results, which the README is careful to label a first baseline with one run per task and setup rather than a verdict, point in three directions. agy suits research best, because DeepSearchQA is the only benchmark where agy cut host spend on both hosts, with Claude Code down 50 percent and Codex down 16 percent. Claude Code's harness suits agy better, because with agy its host spend fell on all three benchmarks between 16 and 50 percent while Codex's rose on two, by 86 percent on SWE-bench and 22 percent on QnA, and the stated reason is mechanical: Claude Code waits in the background until the agy job ends while Codex polls, and every wake-up carries its full context. The third point is the price, with median time per task rising from 2 to 8 minutes alone to 13 to 26 minutes with agy, and tasks solved matching the solo result or dropping by one. ## A plugin package with generated skills and a checked build The repository is a plugin monorepo with one package manifest and several target directories. The published files list includes skills/, pi-skills/, opencode-skills/, the opencode.mjs entry point, companion/, templates/, the skill generator script and the reference documents, and the package declares itself as a pi package and an opencode plugin with a MIT licence. The interesting engineering detail is that the skill directories are generated rather than hand written. A single script, scripts/generate-skills.mjs, produces them, and it is run with a target argument for pi, opencode or the default. Each target has a matching check mode, and the check script is wired into the pack and publish lifecycle, with prepublishOnly running the skills check and a pack:checked script that checks then packs with scripts ignored. The test setup uses the Node test runner, with scripts for cache, schema and integration tests per target. That structure has a practical effect. A persona is defined once and emitted in the format each harness expects, so adding a persona or changing a prompt does not mean editing four copies by hand, and the check step catches a mismatch between the source and the generated files before anything is published. The repository also carries a bench/ directory with the results and its own readme, an assets/ directory with the design images, and a docs/ folder with the reference material and an install guide written for agents rather than humans. ## Version 0.7.4, released three times in three weeks in September and October The project moves quickly. The releases listed are v0.7.4 on 2026-10-03, with the title noting OpenCode plugin support, v0.7.3 on 2026-09-15 with clearer lead delegation and acceptance defaults, and v0.7.2 on 2026-09-13, whose title records that on Windows the workers survive other jobs' cleanup. The last push was on 2026-10-20 and the repository is not archived, so the code is current even though the listed releases stop at v0.7.4. The licence is MIT, and the project metadata points to the repository, the issues tracker and the release notes. Note the versioning: these are pre-1.0 releases, and a release title like workers survive other jobs' cleanup tells you the project is still working through the rough edges of process management on Windows. The dependency picture is unusually light for a plugin system, and that is by design. There is no model SDK in the manifest at all, because the plugin does not talk to Gemini itself. It shells out to the agy CLI that you already installed and authenticated, which means the plugin inherits whatever access that CLI has, including the account and quota behind it, and it also means the plugin is not the thing holding your credentials. ## Conclusion Adopt agy-staff if you are a Claude Code or Codex user whose bottleneck is the cost of a slow frontier model doing small jobs, and you are willing to accept longer task times in exchange for cheap second opinions, research and reviews. Do not adopt it if your work needs deterministic completion times or if the tokens of the extra host model are already your main cost. Before enabling the implementer persona, read the restrictions note, because the personas run unrestricted by default and the project is explicit that the prompt instructions do not provide permission isolation, so run setup with --restricted for anything untrusted. ## FAQ ### What is agy-staff and what does it do? agy-staff is a plugin set that lets a senior coding agent delegate work to Google's Antigravity CLI, agy, which ships fast Gemini 3.8 Flash access. It provides five personas, staffer, researcher, reviewer, implementer and ask, plus a jobs skill, invoked as /agy: in Claude Code and $agy: in Codex. ### How do I install agy-staff for Claude Code and Codex? Install the Antigravity CLI first and check it with agy --version, then add the plugin. For Claude Code run claude plugin marketplace add keli-wen/agy-staff followed by claude plugin install agy@agy-staff, and for Codex run codex plugin marketplace add with the repository URL followed by codex plugin add agy@agy-staff. ### Does agy-staff restrict what the delegated agent can do to my code? Not by default. The staffer, researcher, reviewer and implementer personas run unrestricted, so agy can inspect the repository, run commands and edit files, and the README states that the prompt instructions do not provide permission isolation. A setup step with a --restricted flag is offered as optional hardening for untrusted input. ### What does the agy-staff bench say about delegating work? In Bench v1, agy cut host spend on both hosts for DeepSearchQA, with Claude Code down 50 percent and Codex down 16 percent, and Claude Code's spend fell on all three benchmarks while Codex's rose on two because Codex polls and carries its full context on every wake-up. Median time per task rose from 2 to 8 minutes alone to 13 to 26 minutes with agy, and the README presents this as a first baseline rather than a verdict. ### Which harnesses does agy-staff support? Claude Code, Codex, Pi and OpenCode. The OpenCode adapter is verified on version 1.18.34, the package declares an engine range of greater than or equal to 1.18.34 and below 2, and OpenCode V2 is not covered by the adapter. ## Sources - [Issues](https://github.com/keli-wen/agy-staff/issues) - [keli-wen/agy-staff on GitHub](https://github.com/keli-wen/agy-staff) - [License: MIT](https://github.com/keli-wen/agy-staff/blob/master/LICENSE) - [README](https://github.com/keli-wen/agy-staff/blob/master/README.md) - [Releases](https://github.com/keli-wen/agy-staff/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/keli-wen-agy-staff