Gemini Subagents

Gemini Subagents

Overview

A manager/worker pattern for building frontends. The orchestrator (Claude, or any capable agent) owns taste and architecture: it sets the design direction, splits the work, writes precise specs, and reviews + audits every result. Gemini workers do the bulk code generation — fast and cheap. Quality control stays with the manager because fast models drift toward generic ("slop") output.

Two ways to run the workers:

• Mode A — headless Gemini CLI (automated): the orchestrator auto-dispatches gemini -p calls. Scriptable and parallel.
• Mode B — Antigravity handoff (human-in-the-loop, Gemini 3.5 Flash): the orchestrator writes spec files; a human runs Antigravity (which exposes Gemini 3.5 Flash on consumer accounts) to generate; Antigravity writes code into the same workspace; the orchestrator reads + audits it.

Models

ID gotchas (verify — these change over time): in the Gemini CLI, gemini-3.5-flash, gemini-3-flash, and gemini-flash-latest return ModelNotFoundError. The real Gemini 3 Flash ID is gemini-3-flash-preview. Gemini 3.5 Flash is currently reachable only through Antigravity. Confirm any ID:

gemini -m <model-id> -p "ok"   # "ModelNotFoundError" = wrong ID

Google announced the Gemini CLI stops serving Google One / unpaid tiers (June 18, 2026). On consumer accounts, Mode B (Antigravity) becomes the path to the newest models. Re-check current state before relying on Mode A.

Mode A — headless Gemini CLI (automated)

# stdout mode (manager reviews, then writes the file) — SAFEST default
gemini -m gemini-3-flash-preview -p "<spec>. Output ONLY raw code to stdout. Do NOT use tools or write files. No markdown fences, no explanation."

# autonomous mode (Gemini writes files itself) — scope the cwd, auto-accept actions
cd <scoped-folder> && gemini -m gemini-3-flash-preview -y -p "<spec>; write the result to ./path/file.tsx"

• -p headless · -m model · -y auto-accept tool/file actions
• Parallel: dispatch several gemini -p calls as background jobs (one per section), then collect + review all outputs.
• Strip noise lines (true color, ripgrep, Shell cwd was reset) from captured output before parsing.

Mode B — Antigravity handoff (Gemini 3.5 Flash)

Antigravity is an agentic IDE, not a headless CLI, so it can't be auto-dispatched. The orchestrator and Antigravity share the workspace filesystem; a human is the runner.

1. Orchestrator writes a spec to specs/<section>.md (use spec-template.md) — design tokens + bans + exact files baked in.
2. Human opens Antigravity in the project, sets the model to Gemini 3.5 Flash (/model command → pick it, or the model picker), and says: "Implement specs/.md."
3. Antigravity writes the code files into the repo.
4. Human says "done" — the orchestrator reads the files directly (no pasting).
5. Orchestrator reviews + audits, then writes the next spec or a fix-list spec.

The director's job (non-negotiable)

• Lock ONE aesthetic + exact font/color/spacing tokens before any code.
• Decompose the build into independent sections.
• Bake the rules into every spec — workers do NOT see the orchestrator's skills.
• Review every output; reject slop; never ship a worker's first draft unchecked.

Worker spec must include

Locked aesthetic · exact font/color/spacing tokens · explicit bans (no Bootstrap blue #007bff, no inline event handlers, no templated hero+3-cards, no placeholder/TODO code, no lorem ipsum) · exact file path(s) to write · a short acceptance checklist · the output contract (stdout-only, or the file to write).

Common mistakes

Quick reference

• Worker (CLI): gemini -m gemini-3-flash-preview -p "<spec + output contract>"
• Hard part: gemini-2.5-pro. True Gemini 3.5 Flash: Antigravity (Mode B).
• Always: orchestrator locks tokens → workers generate → orchestrator reviews + audits.

Source: taherfarg/claude-antigravity-subagents — distributed by TomeVault.